Datasets:

zai-org
/

RPC-Bench

+# Watching Too Much Television is Good: Self-Supervised Audio-Visual Representation Learning from Movies and TV Shows
+Anonymous Author(s)
+Affiliation
+Address
+email
+# Abstract
+1 The abundance and ease of utilizing sound, along with the fact that auditory clues
+2 reveal so much about what happens in the scene, make the audio-visual space a
+3 perfectly intuitive choice for self-supervised representation learning. However,
+4 the current literature suggests that training on uncurated data yields considerably
+5 poorer representations compared to the curated alternatives collected in supervised
+6 manner, and the gap only narrows when the volume of data significantly increases.
+7 Furthermore, the quality of learned representations is known to be heavily influ
+8 enced by the size and taxonomy of the curated datasets used for self-supervised
+9 training. This begs the question of whether we are celebrating too early on catching
+10 up with supervised learning when our self-supervised efforts still rely almost exclu
+11 sively on curated data. In this paper, we study the efficacy of learning from Movies
+12 and TV Shows as forms of uncurated data for audio-visual self-supervised learning.
+13 We demonstrate that a simple model based on contrastive learning, trained on a
+14 collection of movies and TV shows, not only dramatically outperforms more com
+15 plex methods which are trained on orders of magnitudes larger uncurated datasets,
+16 but also performs very competitively with the state-of-the-art that learns from
+17 large-scale curated data. We identify that audiovisual patterns like the appearance
+18 of the main character or prominent scenes and mise-en-scène which frequently
+19 occur through the whole duration of a movie, lead to an overabundance of easy
+20 negative instances in the contrastive learning formulation. Capitalizing on such
+21 observation, we propose a hierarchical sampling policy, which despite its simplicity,
+22 effectively improves the performance, particularly when learning from TV shows
+23 which naturally face less semantic diversity.
+# 24 1 Introduction
+25 Recently, there has been tremendous progress in self-supervised learning from still images, where the
+26 standard supervised training has been outperformed in a variety of image-related tasks [7, 8, 15, 29].
+27 The appeal of detaching representation learning from human annotations is rooted not only in the
+28 non-trivial challenges of scaling-up the labeling process, but also in the ill-defined task of determining
+29 a proper taxonomy with generalization power and transferability. Both challenges only exacerbate as
+30 we move from images to videos, where the notion of time is involved and the complexity of visual
+31 concepts increases. Simply considering the number of training instances or even the cardinality of
+32 the label set is not sufficient to conclude if one large-scale supervised dataset is more suitable than
+33 another for transfer learning in video classification tasks [20]. That is, the abundance of attention
+34 which video self-supervised learning has lately received is only to be expected. While many research
+35 efforts in this area extend the contributions made initially in the image domain to the video domain,
+36 others, including our work, have explored harnessing additional modalities such as audio or text for
+37 multi-modal self-supervised learning [2, 3, 4, 22, 27, 31, 37, 36, 39].
+38 From the current state-of-the-art one makes two major conclusions. First, the quality of learned
+39 representations, evaluated by fine-tuning on downstream tasks, is heavily influenced by the size and
+40 taxonomy of the pretraining datasets [2, 3, 39]. Second, an uncurated pretraining dataset yields
+41 considerably poorer representations compared to a curated one and the gap only narrows when the
+42 total amount of pretraining data significantly increases [3]. Curated data refers to likes of supervised
+43 large-scale action recognition and audio classification datasets such as Kinetics [6], IG-Kinetics
+44 [12], AudioSet [11], and YouTube-8M [1]. While the human-annotated labels are not accessed for
+45 self-supervised pretraining, videos being trimmed and from a label set of limited cardinality with
+46 biased sampling distribution1 implicitly acts as a sort of supervision. On the other hand, an uncurated
+47 data refers to likes of IG-Random[3], simply a body of unlabeled videos collected blindly with
+48 none of the aforementioned careful human-involvements. That being said, we know that something
+49 as simple as having access to a clean object-centric training data, like Imagenet, can be indirectly
+50 exploited by contrastive self-supervised learning in image domain to obtain additional performance
+51 gain [41] on the downstream tasks which exhibit similar properties. The analogous to it of course are
+52 the well trimmed closed-set curated datasets which are being extensively used in the literature for
+53 video self-supervised pretraining, while downstream evaluations focus on benchmarks with similar
+54 characteristics. Our work aims at comprehensively exploring the efficacy of learning from Movies
+55 and TV Shows, as forms of uncurated data, for audio-visual self-supervised learning.
+56 Many of us can relate to an experience in movie theaters when the sound of the engine, first perceived
+57 by our left ear, is gradually heard more by the right ear as a car moves from the left side of the
+58 screen to the right side. Another example is a scene in which an object, like a helicopter, approaches
+59 the camera from distance and eventually flies over it. In this case, the perceived sound not only
+60 changes in loudness but also transitions from front to back, in concert with the visuals, giving the
+61 audience a more realistic feeling as if they are indeed positioned behind the camera. Besides, with
+62 art being inherently novel, two movies even if they share genres or revolve around similar story
+63 lines often deliver quite different experiences and portray distinct visuals, thanks to the extremely
+64 artist-driven creative process behind such productions. We hypothesize that the aforementioned high
+65 audio fidelity, and inherent semantic diversity characterize long-form content2 as potentially a very
+66 rich source for self-supervised multi-modal representation learning. It is worth emphasizing that in
+67 spirit of uncurated data, we not only blindly sample from a large collection of movies and TV shows
+68 when constructing our pretraining dataset, but also perform ablation studies on the effect of genre
+69 distribution, the closest we have to taxonomy in the curated datasets, confirming that the quality of
+70 learned representations is agnostic with respect to such statistics.
+71 To the best of our knowledge, we are the first to solely rely on uncurated data and study the efficacy of
+72 self-supervised multi-modal representation learning from movies and TV shows. Despite meaningful
+73 domain gap between our pretraining data and the space of downstream tasks, we obtain representations
+74 which are very competitive with those learned from curated datasets. This is particularly important as
+75 we follow a much simpler modeling approach in comparison with the state-of-the art.
+# 76 2 Related Work
+77 Self-supervised learning techniques define pretext tasks, mostly inspired by the natural structures
+78 in the data, in order to generate supervisory signals for training. Despite the plethora of proposed
+79 pretext tasks in the literature, these approaches can be coarsely divided into two groups, namely
+80 pretext learning, and pretext-invariant methods. Approaches which fall in the former bucket, usually
+81 apply a form of transform, randomly drawn from a parametric family, to the input data then optimize
+82 for predicting the parameters of the chosen transformation. Predicting the relative position of image
+83 patches [9], solving jigsaw puzzles [33], estimating artificial rotations [13], colorization [50], context
+84 encoders learned through inpainting [38], and learning by counting scale and split invariant visual
+85 primitives [34], are among many methods which belong to this category. Similar techniques have
+86 been extended from images to videos [10, 21, 24, 25, 30, 46, 48, 49], where in addition to the
+87 spatial context, the temporal domain, and the arrow of time have been heavily exploited. In contrast,
+88 pretext-invariant methods [5, 7, 8, 15, 18, 17, 29, 35, 39, 44] are built on the concept of maximizing
+89 mutual information across augmented versions of a single instance, and are mostly formulated as
+90 contrastive learning. In other words, a pretext is used to generate different views of a single input
+91 for which the learning algorithm aims to maximize the intra-instance similarity, across variety of
+92 transformations. Our work falls within this category, however we function in a multi-modal realm
+93 employing both audio and video.
+94 Earlier works which harnessed audio and video for representation learning, have leveraged audio
+95 visual temporal synchronization [22, 36], correspondence [4], and cross-modal clustering [3, 37]. The
+96 work by Patrick et al.[39] proposes a generalized data transformation in order to unify a variety of
+97 audio-visual self-supervised pretext tasks through a noise contrastive formulation. This work is close
+98 to ours in choice of objective function and data type, yet we employ no augmentation (except modality
+99 projection in the terminology of [39]), and solely focus on capitalizing the advantages of learning
+100 from long-form content. Morgado et al.[31] show that cross-modal discrimination is important for
+101 learning good audio and video representations, something which was also pointed out earlier in a
+102 clustering framework [3]. Beyond that, [31] generalizes the notion of instance-level positive and
+103 negative examples by exploring cross-modal agreement where multiple instances are grouped together
+104 as positives by measuring their similarity in both the video and audio feature spaces. While we also
+105 adopt a cross-modal noise contrastive estimation loss, we stick with the vanilla version, instance-level
+106 positive and negatives, and do not use any memory bank feature representations. Finally, Alayrac et
+107 al.[2] recently proposed a multi-modal versatile network capable of simultaneously learning from
+108 audio, video and text. Building on the intuition that different modalities are of different semantic
+109 granularity, audio and video are first compared in a fine-grained space while text is compared with
+110 the aforementioned modalities in a lower dimensional coarse-grained space. In our experiments, we
+111 compare with a variant of [2] where only audio and video modalities are utilized.
+# 112 3 Approach
+113 Notations and Architecture. Our pretraining dataset is denoted by $\mathcal { X } \ : = \ : \{ \mathcal { X } _ { n } | n \in [ 1 \cdot \cdot \cdot N ] \}$
+114 where $\mathcal { X } _ { n } = \{ x _ { n , m } | m \in [ 1 \cdots M _ { n } ] \}$ contains $M _ { n }$ non-overlapping audiovisual snippets which are
+115 temporally segmented from the duration of the $n ^ { t h }$ long-form content in the dataset. Each snippet
+116 includes both audio and video modalities, formally $x _ { n , m } = ( a _ { n , m } , v _ { n , m } )$ , where $a _ { n , m } \in \mathbb { R } ^ { 1 \times \pmb { \tilde { P } } \times Q }$
+117 and $v _ { n , m } \in \mathbb { R } ^ { 3 \times T \times H \times W }$ . $T , H$ , and $W$ denote the number of frames, height and width of the video,
+118 while $P$ , and $Q$ respectively stand for the number of mel filters, and audio frames. Video and audio are
+119 processed through 18-layers deep $\mathrm { R } ( 2 { + } 1 ) \mathrm { D }$ [45] and ResNet [16] architectures, respectively referred
+120 to as $f : \mathbb { R } ^ { 3 }  \check { \mathbb { R } } ^ { d _ { f } }$ and $\dot { \boldsymbol { g } } : \mathbb { R } ^ { 1 } \dot {  } \mathbb { R } ^ { d _ { g } }$ . Inspired by [7], we use projection heads, $\boldsymbol { h } _ { f } : \mathbb { R } ^ { d _ { f } } \overset { \cdot } {  } \mathbb { R } ^ { d }$ and
+121 $h _ { g } : \mathbb { R } ^ { d _ { g } }  \mathbb { R } ^ { d }$ , to map corresponding representations into a common $d$ -dimensional space before
+122 computing the contrastive loss. The shallow architecture of $h _ { f }$ and $h _ { g }$ consists of two convolution
+123 layers, separated by Batch Normalization [19] and ReLU [32], followed by global average pooling.
+124 Once self-supervised pretraining finished, we discard the projection heads and fine-tune $f$ and $g$ for
+125 respective downstream tasks.
+126 Loss Function. With a slight abuse of notation3, $\boldsymbol { B } = \{ x _ { i } = ( a _ { i } , v _ { i } ) | i \in [ 1 \cdot \cdot \cdot B ] \}$ represents a
+127 minibatch of size $B$ , where video and audio modalities associated with the $i ^ { t h }$ sample, $x _ { i }$ , are denoted
+128 by $v _ { i }$ and $a _ { i }$ . We use $z _ { v } ^ { i } = h _ { f } ( f ( v _ { i } ) )$ and $z _ { a } ^ { i } = h _ { g } ( g ( a _ { i } ) )$ to represent the associated embeddings
+129 generated by projection heads, and optimize the noise-contrastive loss [14] shown in 1 in order
+130 to maximize the symmetric joint probability between audio and video. For the $i ^ { t h }$ element in the
+131 minibatch, $( z _ { v } ^ { i } , z _ { a } ^ { i } )$ serves as the positive pair, while assuming negative pairs for both modalities,
+132 $\mathcal { N } _ { i } = \{ ( z _ { v } ^ { i } , z _ { a } ^ { j } ) , ( \stackrel { \sim } { z } _ { v } ^ { j } , z _ { a } ^ { i } ) | j \in [ 1 \cdots B ] , i \neq j \}$ constitutes the set of negative pairs.
+$$
+\mathcal { L } = - \sum _ { i = 1 } ^ { B } \log \left( \frac { e ^ { ( z _ { v } ^ { i } ) ^ { \top } ( z _ { a } ^ { i } ) } } { e ^ { ( z _ { v } ^ { i } ) ^ { \top } ( z _ { a } ^ { i } ) } + \displaystyle \sum _ { ( z _ { v } ^ { \prime } , z _ { a } ^ { \prime } ) \in \mathcal { N } _ { i } } e ^ { ( z _ { v } ^ { \prime } ) ^ { \top } ( z _ { a } ^ { \prime } ) } } \right)
+$$
+133 Most of the previous works [2, 31, 39] normalize the embeddings before computing the contrastive
+134 loss and employ a temperature hyper-parameter, often denoted by $\tau$ as in [2, 31], to control the
+35 smoothness for the distribution of pairwise similarities. In contrast, we have chosen to operate in an
+36 unnormalized embedding space. Besides the obvious benefit of eliminating the need for tuning $\tau$ , we
+7 empirically show that such decision does not affect the quality of the learned representations.
+138 Sampling Policy. Contrastive loss function shown in Equation 1 is computed over $B$ training
+139 instances, each in form of an audiovisual snippet. A naive sampling policy may ignore the fact that
+140 snippets comprising the pretraining dataset are in fact temporal segments that were trimmed from
+141 longer-form contents, i.e. movies and TV shows. Such an assumption treats our training data as
+142 independent and identically distributed random variables from $\textstyle \bigcup _ { n = 1 } ^ { N } { \mathcal { X } } _ { n }$ , which constitutes the default
+143 sampling policy that is commonly used in the general deep learning literature. However, in reality,
+144 commonalities and correlations do exist along the temporal axis of a movie or TV show, things like
+145 audio mastering artifacts, frequent appearance of the main character’s face and voice, thematic music,
+146 repetitive scenes and mise-en-scène4, all of which contribute to breaking the previously discussed
+147 i.i.d assumption. This is even more pronounced when we deal with multiple episodes of the same TV
+148 show appearing in the pretraining dataset5. Note that, sampling from no video data is going to be i.i.d
+149 but in this case the temporal correlations extend for much longer given our entities are movies and TV
+150 shows. Thus, it is more accurate to think of $\mathcal { X }$ having multiple underlying domains, oriented towards
+151 exclusive properties which different long-form contents are characterized by. We hypothesize that
+152 during training, model gradually discovers such patterns of commonalities, which are not semantically
+153 valuable, and latches onto those to quickly minimize Equation 1 leading to poor generalization6. The
+154 reason being $B \ll N$ , hence for $n \sim \mathbb { U } ( 1 , N )$ and $m \neq m ^ { \prime }$ , $\mathsf { P } ( x _ { n , m } \in B \land x _ { n , m ^ { \prime } } \in B )$ is negligible.
+155 In other words, the set of negative pairs in Equation 1 mainly includes pairs for which audio and
+156 video come from two different movies or TV shows, thus due to the aforementioned artifacts behave
+157 as easy negatives.
+158 In order to quantitatively measure our hypothesis, we define different distributions, shown in Equation
+159 2, over the space of audio-visual similarity. $S ^ { + }$ indicates the space of correct matches, i.e. where
+160 audio and video correspond to the same snippet. $S ^ { - }$ indicates the space where audio and video do
+161 not correspond yet belong to the same movie or TV show. Finally, $\bar { \boldsymbol { S } } ^ { \neq }$ indicates the space in which
+162 audio and video are sampled from two distinct long-form content, hence naturally do not correspond.
+$$
+( z _ { v } ^ { n , m } ) ^ { \mathsf { T } } ( z _ { a } ^ { n ^ { \prime } , m ^ { \prime } } ) \sim \left\{ { \begin{array} { l l } { S ^ { + } , } & { { \mathrm { i f } } \ n = n ^ { \prime } \wedge m = m ^ { \prime } } \\ { S ^ { - } , } & { { \mathrm { i f } } \ n = n ^ { \prime } \wedge m \neq m ^ { \prime } } \\ { S ^ { \neq } , } & { { \mathrm { i f } } \ n \neq n ^ { \prime } \wedge \forall ( m , m ^ { \prime } ) } \end{array} } \right.
+$$
+163 With that, and $\mathsf { K L }$ denoting Kullback–Leibler divergence, $\mathsf { K L } ( S ^ { - } \parallel S ^ { + } )$ measures the expected
+164 difference between positive and negative pairs within the same movie or TV show. Ideally, this should
+165 increase as the training progresses, since the model gradually learns audio-video correspondence by
+166 minimizing Equation 1. Meanwhile, the i.i.d assumption suggests $\mathsf { K L } ( \mathcal { S } ^ { - } \parallel \mathcal { S } ^ { + } ) \simeq \mathsf { K L } ( \mathcal { S } ^ { \neq } \parallel \mathcal { S } ^ { + } )$
+167 and $\mathsf { K L } ( \mathcal { S } ^ { - } \parallel \mathcal { S } ^ { \neq } ) \simeq \boldsymbol { 0 }$ , yet as we empirically illustrate later, $\mathsf { K L } ( \mathcal { S } ^ { - } \parallel \mathcal { S } ^ { + } ) < \mathsf { K L } ( \mathcal { S } ^ { \neq } \parallel \mathcal { S } ^ { + } )$ and
+168 ${ \mathsf { K L } } ( S ^ { - } \parallel S ^ { \neq } )$ is rather large, indicating that, upon convergence and on a held-out set, model has
+169 a harder time pushing apart negative pairs when audio and video come from the same underlying
+170 long-form content. Next, we explain how a simple alternative policy which samples $k$ snippets
+171 from each long-form content effectively reduces both of the discrepancy measures, referring to
+172 $\mathsf { K L } ( S ^ { - } \parallel S ^ { \neq } )$ and $\mathsf { K L } ( \mathcal { S } ^ { \neq } \parallel \mathcal { S } ^ { + } ) - \mathsf { K L } ( \mathcal { S } ^ { - } \parallel \mathcal { S } ^ { + } )$ , while yielding better generalization on a range
+173 of downstream tasks.
+174 To ameliorate the aforementioned optimization challenge, we take a hierarchical approach. In
+175 particular, we first uniformly sample a long-form content, $n \sim \mathbb { U } ( 1 , N )$ , and then draw $k$ distinct
+176 snippets from ${ \mathcal { X } } _ { n }$ , creating $\{ { \bar { x } } _ { n , m } | { \bar { m } } \in { \mathcal { M } } _ { n } \}$ , where $\mathcal { M } _ { n } \subset [ 1 \cdots M _ { n } ]$ and $| { \mathcal { M } } _ { n } | = k$ . This ensures
+177 that for $x _ { i } \in B$ , ${ \mathcal { N } } _ { i }$ always includes $2 k - 2$ pairs sampled from the same movie or TV show to
+178 which $x _ { i }$ belongs. By putting constraints on $\mathcal { M } _ { n }$ , specifically how temporally far from each other
+179 the $k$ samples are drawn, we may go one step further and to some extent control the audiovisual
+180 similarity between snippets. This serves as an additional nob to tune for hard negative sampling.
+181 The intuition is that, the larger narrative of a professionally made movie or TV show is composed of
+182 shorter units called scene. Each scene comprises a complete event, action, or block of storytelling and
+183 normally takes place in one location and deals with one action. That is, if our samples are temporally
+184 close, it is more likely for corresponding snippets to be highly correlated and/or look/sound alike.
+185 $k \leq \mathrm { m a x } [ \mathcal { M } _ { n } ] - \mathrm { m i n } [ \mathcal { M } _ { n } ] + 1 \ \bar { \leq } \ w \leq \bar { M } _ { n }$ defines the bounds on our sampling policy, where $w$
+186 standing for a sampling window, determines the farthest two out of $k$ samples drawn from ${ \mathcal { X } } _ { n }$ can
+187 be. Accordingly, $w = k$ represents the case where all $k$ samples are temporally adjacent, hence
+188 the expected audiovisual similarity is maximized due to temporal continuity in content. We show
+189 that having such level of hard negatives, even with a small $k$ , prevents proper training and results in
+190 performance degradation. On the other hand, $w = M _ { n }$ indicates random sampling where no temporal
+191 constraint is imposed on $\mathcal { M } _ { n }$ , thus samples are less likely to be drawn from adjacent time-stamps.
+192 In this case, expected audiovisual similarity (i.e. hardness of negative pairs) is mainly derived from
+193 global content-exclusive artifacts like, color palette, frequent appearance of the main character’s face
+194 and voice, repetitive scenes, and etc. The rest of the spectrum provides middle grounds where two
+195 samples drawn from ${ \mathcal { X } } _ { n }$ can at most be $w + 1$ snippets apart, something reminiscent of temporal
+196 locality. Our sampling policy can be easily implemented in a few lines of Python. Please refer to
+197 supplemental material for further details.
+# 98 4 Experiments
+# 4.1 Experimental Setup
+Datasets and Reproducibility. We use full-length movies and episodes of TV shows for selfsupervised pretraining. Titles are randomly chosen from a large collection spanning over a variety of genres, namely Drama, Comedy, Action, Horror, Thriller, Sci-Fi and Romance. All audio is in English language. Our Movie dataset, consists of 3.6K films with an average duration of 105 minutes. Our TV dataset includes 9.2K episodes from a total of 581 shows with an average duration of 42 minutes per episode. Each of our datasets comprises 0.7 years worth of uncurated audiovisual content, which is significantly smaller than IG-Random [3] with variants at 5 and 21 years. Scaling up our pretraining datasets to volumes comparable to the IG-Random [3] while possible is non-trivial and demands dramatically larger compute resources for training, something which we currently cannot afford. Given that we cannot publicly release our dataset due to copyright reasons, we acknowledge that it is not possible for other research groups to fully reproduce our results. However, we intend to make available the pretrained models and hope that research community finds them, along with the other contributions of this work, of value whether within the context of self-supervised learning or adoption for various downstream tasks. We would like to emphasize that similar limitations have precedents in multiple earlier works including but not limited to [3, 12, 26, 43]. To evaluate the efficacy of self-supervised audio-visual representation learning from movies and TV shows, we follow recent works [3, 39, 31, 2] and benchmark UCF101[42] and HMDB51[23] for action recognition, along with ESC50[40] for audio classification. Results for the ablation studies are reported on the split-1 of the corresponding datasets. Following the standard protocol, we report the average performance over all splits when we are comparing with the state-of-the-art.
+Pretraining. Unless mentioned otherwise, we use video snippets with 16 frames at 5 fps. For data augmentation, we resize the shorter side to 190 pixels, then randomly crop them into $1 5 8 \times 1 5 8$ pixels. As for sound, we compute mel spectrogram from the raw audio at 48K sample rate using $9 6 ~ \mathrm { m e l }$ filters and an FFT window of 2048, while the number of samples between successive frames is set to 512. For data augmentation, we randomly drop out up to $2 5 \%$ from either temporal or frequency axis of the 2-D mel spectrogram image. Training uses a batch size of 512 and takes on average 42 hours on 8 NVIDIA A100 GPUs. The dimension of audio-video joint embedding space, $d$ , is set to 512.
+Downstream Evaluation. For training on UCF101 [42] and HMDB51 [23], we use video clips that are 32 frames long at 10 fps. Unless mentioned otherwise, these clips are randomly chosen from the duration of the video instances. A scale jittering range of [181, 226] pixels is used and we randomly crop the video into $1 5 8 \times 1 5 8$ pixels. Furthermore, random horizontal flipping and color jittering are employed. During inference, 10 temporal clips are uniformly sampled where each is spatially cropped in 3 ways (left, center, right) resulting in a total of 30 views. We then average the model predictions across these 30 views and report top-1 classification accuracy. For training on ESC50 [40], we use 3-seconds clips which are randomly chosen from the duration of the audio instances and apply time and frequency masking to spectrogram images for data augmentation. The maximum possible length of the mask is $50 \%$ of the corresponding axis. We do not use any scale jittering or random cropping on the spectrograms. During inference, 10 temporal clips are uniformly sampled
+238 and we average the model predictions across these 10 views and report top-1 classification accuracy.
+239 For further implementation details, please refer to the supplemental material.
+![](images/fe3fa73a08dcc0cf4d8abc90b46bc44d9e947f10dd7a4be6b2c5ef8be9fc6fa5.jpg)
+Figure 1: Ablation study of the proposed sampling policy on reducing the discrepancy measures.
+# 240 4.2 Ablation Study
+In the following, we discuss multiple ablation studies to assess our main hypothesis that, a hierarchical sampling policy, as described in Section 3, enables better representations to be learned by increasing the portion of hard negative pairs which the contrastive loss function observes. Here, pretraining uses $90 \%$ of either Movie or TV dataset, while the remaining $10 \%$ constitute a held-out validation set7 on which we report the discrepancy measures.
+Sample size $( k )$ Figure 1a illustrates that compared to the baseline sampling denoted by $k = 1$ , our approach $k > 1 ,$ ) effectively shrinks the gap between $S ^ { - }$ and ${ \mathcal { S } } ^ { \neq }$ when measured either directly or against $S ^ { + }$ . Its pattern of behavior also perfectly follows our earlier intuition (ref. Section 3). In particular, given a fixed minibatch budget, a larger $k$ favors more training instances to be sampled from fewer number of long-form contents. That increases the portion of hard negative pairs, thus pushes the contrastive loss to more aggressively separate mismatched audio-video pairs from the same movie, which leads model to maintain less of the content-exclusive artifacts in the embedding space. In the most extreme case, $k = 6 4$ , all the training instances are sampled from the same movie. From Table 1, we observe that different variants of our sampling policy, with no imposed temporal constraint, i.e. $w = M _ { n }$ , outperform the baseline on all three downstream tasks
+![](images/bf6588579202158118301afb6941dec79d9e5cdccd64dae9b2636f3ddcd67c90.jpg)
+Figure 2: Effect of color jitter on the discrepancy measures.
+Sampling window $\mathbf { \Pi } ^ { ( w ) }$ Smaller $w$ forces samples that belong to same movie to be drawn from a shorter temporal window, hence growing the probability that they look/sound very much alike (i.e. harder negative pairs). That is, it should further diminish the discrepancy measures. Figure 1b illustrates this behavior where we gradually increase $w$ while $k = 1 6$ . However, from Table 1, it does not seem that tuning for $w$ , i.e $w \neq M _ { n }$ , provides a meaningful gain on downstream tasks. This implies that commonalities which persist throughout the duration of a movie are sufficiently powerful signals to be exploited for generating hard negatives. We hypothesize that different scenes both within and across different movies and TV shows are of variety of length, thus a fixed $w$ is sub-optimal. Ideally, we should identify scene boundaries and and dynamically modify $w$ during sampling, something which we leave for future iterations of this work.
+Temporally adjacent samples. Along the lines of previous observations, Figure 1c shows that indeed drawing temporally adjacent snippets from the same long-form content, i.e. $w = k$ , results in aggressively reducing the discrepancy measures. This behavior is agnostic with respect to $k$ yet exacerbates as $k$ grows. Note that, the contrastive loss is an instance-discrimination objective function. Therefore, forcing it to distinguish between temporally adjacent snippets, that naturally sound and look extremely similar, leaves no choice for the model but to discard valuable semantic notions, which predictably leads to poor representations, also confirmed by result reported in Table 1.
+Table 1: Ablation study of the proposed sampling policy on different downstream tasks, measured by top-1 classification accuracy.
+<table><tr><td colspan="5">pretraining dataset: Movie</td></tr><tr><td>k</td><td>w</td><td>HMDB51</td><td>ESC50</td><td>UCF101</td></tr><tr><td>1</td><td>1</td><td>60.32</td><td>86.50</td><td>85.69</td></tr><tr><td>4</td><td>Mn</td><td>61.37</td><td>89.91</td><td>85.38</td></tr><tr><td>8</td><td>Mn</td><td>62.09</td><td>88.75</td><td>86.06</td></tr><tr><td>16</td><td>Mn</td><td>62.92</td><td>88.33</td><td>86.30</td></tr><tr><td>32</td><td>Mn</td><td>61.04</td><td>88.00</td><td>85.98</td></tr><tr><td>64</td><td>Mn</td><td>61.30</td><td>86.83</td><td>85.43</td></tr><tr><td>16</td><td>64</td><td>60.26</td><td>87.00</td><td>83.61</td></tr><tr><td>16</td><td>128</td><td>60.58</td><td>86.50</td><td>85.30</td></tr><tr><td>16</td><td>256</td><td>62.02</td><td>87.75</td><td>84.85</td></tr><tr><td>16</td><td>512</td><td>61.30</td><td>87.08</td><td>85.38</td></tr><tr><td>16</td><td>1024</td><td>60.65</td><td>86.16</td><td>84.61</td></tr><tr><td>16</td><td>2048</td><td>61.83</td><td>87.66</td><td>85.11</td></tr><tr><td>4</td><td>4</td><td>60.19</td><td>88.00</td><td>84.66</td></tr><tr><td>16</td><td>16</td><td>56.86</td><td>88.75</td><td>82.71</td></tr><tr><td>64</td><td>64</td><td>57.45</td><td>84.58</td><td>82.68</td></tr><tr><td colspan="5">pretraining dataset:TV</td></tr><tr><td>k</td><td>w</td><td>HMDB51</td><td>ESC50</td><td>UCF101</td></tr><tr><td>1</td><td>1</td><td>56.40</td><td>85.50</td><td>84.37</td></tr><tr><td>8</td><td>Mn</td><td>61.50</td><td>87.50</td><td>85.96</td></tr><tr><td>16</td><td>Mn</td><td>61.69</td><td>89.00</td><td>85.64</td></tr><tr><td>8</td><td>64</td><td>60.58</td><td>88.00</td><td>85.96</td></tr><tr><td>8</td><td>128</td><td>60.00</td><td>85.66</td><td>85.77</td></tr><tr><td>16</td><td>256</td><td>61.30</td><td>86.41</td><td>85.01</td></tr></table>
+Movies vs. TV Shows. To confirm that our sampling policy behaves consistently across both movies and TV shows, Figure 1d illustrates the discrepancy measures computed on TV dataset. We observe similar effectiveness when using $k$ and $w$ as tuning nobs for reducing either ${ \mathsf { K L } } ( S ^ { - } \parallel S ^ { \neq } )$ or the gap between ${ \mathsf { K L } } ( S ^ { - } \parallel S ^ { + } )$ and $\mathsf { K L } ( S ^ { \neq } \parallel S ^ { + } )$ . Table 1 demonstrates that different variants of our approach significantly outperform the baseline, i.e. $k = 1$ . We attribute the larger gains achieved when using TV instead of Movie dataset to the fact that content diversity is naturally lower when pretraining on TV shows since each one includes many episodes that all are characterized with the same content-exclusive artifacts.
+Color jitter. We have established so far that commonalities which persist throughout the duration of a long-form content, things likely associated with color pallet, frequent appearance of the main character’s face and voice, and repetitive scenes can be exploited for learning better representations. That is, one may naturally assume that employing data augmentation techniques like color jitter should be helpful since by distorting content-exclusive visual artifacts, color jitter is expected to reduce ${ \mathsf { K L } } ( S ^ { - } \parallel S ^ { \neq } )$ . Figure 2 illustrates the effect of color jitter, where brightness, contrast, and saturation jitter values are chosen uniformly from $\mathtt { [ m a x ( 0 , 1 - \sigma ) , 1 + \sigma ] }$ . We observe that color jitter reduces the discrepancy measures for the baseline but not as much as it can be obtained by our proposed sampling policy $( k > 1 )$ ), and even then according to Table 2 only yields a slight gain on downstream tasks.
+$\ell _ { 2 }$ -normalized feature space. The common practice [2, 31, 39, 7] is to compute contrastive loss in $\ell _ { 2 }$ -normalized feature space, where according to [47] the temperature hyper-parameter, $\tau$ , controls the strength of penalties on hard negative samples. We explored this with two widely-used $\tau$ values. From Table 2, we observe that compared to operating in an unnormalized embedding space, adopting such design choice results in a large performance drop on HMDB51[23] while other downstream benchmarks see only negligible gains.
+Table 4: Effect of self-supervised learning from curated versus uncurated data on different downstream tasks. The “years” column indicates the duration of the pretraining datasets in years.
+<table><tr><td>method</td><td>pretraining dataset</td><td>uncurated</td><td>years</td><td>HMDB51</td><td>ESC50</td><td>UCF101</td></tr><tr><td>Ours</td><td>Movie</td><td></td><td>0.7</td><td>62.9</td><td>88.3</td><td>86.3</td></tr><tr><td>Ours</td><td>TV</td><td></td><td>0.7</td><td>61.7</td><td>89.0</td><td>85.6</td></tr><tr><td>XDC[3]</td><td>IG-Random16M</td><td>√</td><td>5</td><td>55.2</td><td>84.3</td><td>84.1</td></tr><tr><td>XDC[3]</td><td>IG-Random65M</td><td>√</td><td>21</td><td>61.2</td><td>86.3</td><td>88.8</td></tr><tr><td>XDC[3]</td><td>IG-Kinetics16M</td><td>×</td><td>5</td><td>57.3</td><td>82.5</td><td>87.6</td></tr><tr><td>XDC[3]</td><td>IG-Kinetics65M</td><td>×</td><td>21</td><td>63.1</td><td>84.8</td><td>91.5</td></tr></table>
+Curated vs. Uncurated data. To the best of our knowledge, the only other uncurated dataset used for audio-visual self-supervised learning is IGRandom[3]8. Table 4 confirms that learning from uncurated movies and TV shows is extremely effective. Our results significantly exceed those of XDC[3] obtained on IG-Random16M despite using a simpler model and 7 times smaller volume of pretraining data. Even in comparison to IG-Random65M with 30 times larger data, we obtain better performances on 2 out of 3 benchmarks. The most promising of our findings though is how competitive our results are against XDC[3] when it is trained on variants of IG-Kinetics which are not only curated but also orders of magnitude larger. With all that, we confidently reject the notion that audio-visual self-supervised learning from uncurated data considerably lags behind utilizing large-scale curated datasets.
+Effect of genre. The distribution of genre among movies used in our pretraining is the closest we have to taxonomy in the curated datasets. So, it is worth examining the quality of our learned representations under various genre distributions. To do so, given a fixed pretraining budget $( N = 1 . 6 \mathsf { K } )$ , we compare four different scenarios where movies used in the pretraining are distributed i) non-uniformly over all genres except Drama, and Comedies, ii) non-uniformly over Drama, and Comedies, iii) uniformly over all genres, and iv) non-uniformly over all genres. Table 3 confirms that indeed there is very little difference between the aforementioned setups when it comes to transfer learning to the downstream tasks.
+Table 2: Effect of color jitter $( \sigma )$ and computing contrastive loss in $\ell _ { 2 }$ -normalized embedding space with temperature hyper-parameter $( \tau )$ on different downstream tasks.
+<table><tr><td>k</td><td>0</td><td>HMDB51</td><td>ESC50</td><td>UCF101</td></tr><tr><td>1</td><td>0.0</td><td>60.32</td><td>86.50</td><td>85.69</td></tr><tr><td>16</td><td>0.0</td><td>62.92</td><td>88.33</td><td>86.30</td></tr><tr><td>1</td><td>1.0</td><td>60.45</td><td>87.66</td><td>84.82</td></tr><tr><td>16</td><td>0.5</td><td>60.13</td><td>87.75</td><td>85.98</td></tr><tr><td>16</td><td>1.0</td><td>61.11</td><td>88.33</td><td>85.93</td></tr><tr><td>k</td><td>T</td><td>HMDB51</td><td>ESC50</td><td>UCF101</td></tr><tr><td>16</td><td>0.07</td><td>60.78</td><td>87.08</td><td>86.86</td></tr><tr><td>16</td><td>0.30</td><td>60.78</td><td>89.25</td><td>85.72</td></tr></table>
+Table 3: Effect of genre distribution in Movie dataset on different downstream tasks. Experiments are conducted with input spatial resolution of $1 1 2 \times 1 1 2$ pixels.
+<table><tr><td>setting</td><td>HMDB51</td><td>ESC50</td><td>UCF101</td></tr><tr><td>i</td><td>57.58</td><td>86.50</td><td>82.44</td></tr><tr><td>ii</td><td>56.99</td><td>85.50</td><td>82.39</td></tr><tr><td>i</td><td>56.27</td><td>85.25</td><td>82.87</td></tr><tr><td>iv</td><td>56.40</td><td>86.75</td><td>83.24</td></tr></table>
+# 4.3 Comparison with state-of-the-art
+Table 5 compares our proposed approach of learning from Movies and TV shows against the best performing audio-visual self-supervised learning methods. In general, our numbers are comparable with the best existing results reported in the literature, even with much less data and considerably simpler model/training procedure9. It is interesting that training on Movie dataset alone obtains
+Table 5: Comparison with state-of-the-art. Dataset abbreviations: AudioSet[11], HowTo100M[28], IG-Kinetics65M [12]; their length in years is given in the “years” column. “Arch.” denotes the architecture of video backbone $( f )$ . [2]† indicates when the corresponding model use only audio and video, and not text modality. For a fair comparison, when using only Movie dataset, we train for twice as many epochs as our other variants in order to match their total number of gradient updates.
+<table><tr><td>Method</td><td>Arch.</td><td>pretraining dataset</td><td>curated</td><td>years</td><td>HMDB51</td><td>UCF101</td><td>ESC50</td></tr><tr><td>GDT[39]</td><td>R(2+1)D-18</td><td>AS</td><td></td><td>1</td><td>66.1</td><td>92.5</td><td>88.5</td></tr><tr><td>GDT[39]</td><td>R(2+1)D-18</td><td>IG65M</td><td></td><td>21</td><td>72.8</td><td>95.2</td><td></td></tr><tr><td>XDC[3]</td><td>R(2+1)D-18</td><td>AS</td><td>√</td><td>1</td><td>61.0</td><td>91.2</td><td>84.8</td></tr><tr><td>XDC[3]</td><td>R(2+1)D-18</td><td>IG65M</td><td>√</td><td>21</td><td>67.4</td><td>94.2</td><td></td></tr><tr><td>AVTS[22]</td><td>MC3</td><td>AS</td><td></td><td>1</td><td>61.6</td><td>89.0</td><td>82.3</td></tr><tr><td>AVID[31]</td><td>R(2+1)D-18</td><td>AS</td><td></td><td>1</td><td>64.7</td><td>91.5</td><td>89.1</td></tr><tr><td>MMV[2]+</td><td>R(2+1)D-18</td><td>AS</td><td>√</td><td>1</td><td>70.1</td><td>91.5</td><td>85.6</td></tr><tr><td>MMV[2]+</td><td>S3D-G</td><td>AS</td><td>√</td><td>1</td><td>68.2</td><td>90.1</td><td>86.1</td></tr><tr><td>MMV[2]+</td><td>S3D-G</td><td>AS+HT</td><td>√</td><td>16</td><td>68.3</td><td>91.1</td><td>87.2</td></tr><tr><td>Ours (k=16)</td><td>R(2+1)D-18</td><td>Movie</td><td>X</td><td>0.7</td><td>64.5</td><td>87.9</td><td>88.8</td></tr><tr><td>Ours (k=8)</td><td>R(2+1)D-18</td><td>Movie+TV</td><td>X</td><td>1.4</td><td>65.0</td><td>87.7</td><td>89.1</td></tr><tr><td>Ours (k=16)</td><td>R(2+1)D-18</td><td>Movie+TV</td><td>X</td><td>1.4</td><td>65.1</td><td>88.5</td><td>89.1</td></tr><tr><td>Ours (k=32)</td><td>R(2+1)D-18</td><td>Movie+TV</td><td>X</td><td>1.4</td><td>65.6</td><td>88.7</td><td>88.2</td></tr></table>
+358 comparable performance to the cases where both TV and Movie datasets are used for pretraining.
+359 This further confirms the richness of the training data which movies and TV shows can provide
+360 to self-supervised learning problems. We also see that increasing $k$ even beyond 8 gives further
+361 incremental gains on action recognition benchmarks.
+# 362 5 Conclusion
+363
+364
+365
+366
+367
+368
+369
+370
+371
+372
+373
+Despite its amazing recent progress, state-of-the-art self-supervised learning still heavily relies on supervised, i.e. curated, large-scale datasets for pretraining. In this work, we have shown that pretraining solely on uncurated data in forms of movies and TV shows, even at a comparatively small scale, can give rise to representations which are capable of competing with the state-of-theart of more complex architectures trained on larger curated datasets. This comes contrary to the current literature which tends to suggest that learning from uncurated data largely falls behind the use of curated alternatives. We intentionally made design decisions to keep our approach and training strategy as simple as possible to demonstrate that learning decently powerful audio-visual representations does not necessarily require gigantic data and compute resources. Through extensive set of experiments, our work establishes for the first time the efficacy of self-supervised learning of audio-visual representations from movies and TV shows.
+# 374 6 Broader impact
+Potential benefits. Our work shows that competitive multimodal representations can be learned from a comparatively small volume of uncurated data in the form of movies and TV shows. Besides minimizing any sort of human-involvement, which we believe must have already been paid an extra attention to in the literature, our work demonstrates that one does not require gigantic data and compute resources for effective self-supervised pretraining. Such results promise a more democratized research arena where smaller groups are not alienated due lack of sufficient compute resources. More importantly, lowering the compute requirements naturally reduces any environmental effects which training these models can potentially have.
+Potential risks. Any machine learning method is susceptible to the potential underlying biases in the data. This is more important for self-supervised methods that deal with huge volumes, often not evaluated by diverse group of humans for any fairness concerns. The same is generally true in our case which requires us to make sure that titles that are included in training are diverse and inclusive.
+387 References
+388 [1] S. Abu-El-Haija, N. Kothari, J. Lee, P. Natsev, G. Toderici, B. Varadarajan, and S. Vijayanarasimhan.
+389 Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016.
+390 [2] J.-B. Alayrac, A. Recasens, R. Schneider, R. Arandjelovic, J. Ramapuram, J. De Fauw, L. Smaira, ´
+391 S. Dieleman, and A. Zisserman. Self-Supervised MultiModal Versatile Networks. In NeurIPS, 2020.
+392 [3] H. Alwassel, D. Mahajan, B. Korbar, L. Torresani, B. Ghanem, and D. Tran. Self-supervised learning by
+393 cross-modal audio-video clustering. In Advances in Neural Information Processing Systems, volume 33,
+394 pages 9758–9770, 2020.
+395 [4] R. Arandjelovic and A. Zisserman. Look, listen and learn. In Proceedings of the IEEE International
+396 Conference on Computer Vision, pages 609–617, 2017.
+397 [5] P. Bachman, R. D. Hjelm, and W. Buchwalter. Learning representations by maximizing mutual information
+398 across views. In Advances in Neural Information Processing Systems, pages 15535–15545, 2019.
+399 [6] J. Carreira and A. Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In
+400 proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299–6308,
+401 2017.
+402 [7] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual
+403 representations. arXiv preprint arXiv:2002.05709, 2020.
+404 [8] T. Chen, S. Kornblith, K. Swersky, M. Norouzi, and G. Hinton. Big self-supervised models are strong
+405 semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020.
+406 [9] C. Doersch, A. Gupta, and A. A. Efros. Unsupervised visual representation learning by context prediction.
+407 In Proceedings of the IEEE international conference on computer vision, pages 1422–1430, 2015.
+408 [10] B. Fernando, H. Bilen, E. Gavves, and S. Gould. Self-supervised video representation learning with
+409 odd-one-out networks. In Proceedings of the IEEE conference on computer vision and pattern recognition,
+410 pages 3636–3645, 2017.
+411 [11] J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter.
+412 Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE International Conference
+413 on Acoustics, Speech and Signal Processing (ICASSP), pages 776–780. IEEE, 2017.
+414 [12] D. Ghadiyaram, D. Tran, and D. Mahajan. Large-scale weakly-supervised pre-training for video action
+415 recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages
+416 12046–12055, 2019.
+417 [13] S. Gidaris, P. Singh, and N. Komodakis. Unsupervised representation learning by predicting image
+418 rotations. arXiv preprint arXiv:1803.07728, 2018.
+419 [14] M. Gutmann and A. Hyvärinen. Noise-contrastive estimation: A new estimation principle for unnormalized
+420 statistical models. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and
+421 Statistics, pages 297–304, 2010.
+422 [15] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick. Momentum contrast for unsupervised visual representation
+423 learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages
+424 9729–9738, 2020.
+425 [16] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the
+426 IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.
+427 [17] O. J. Hénaff, A. Srinivas, J. De Fauw, A. Razavi, C. Doersch, S. Eslami, and A. v. d. Oord. Data-efficient
+428 image recognition with contrastive predictive coding. arXiv preprint arXiv:1905.09272, 2019.
+429 [18] R. D. Hjelm, A. Fedorov, S. Lavoie-Marchildon, K. Grewal, P. Bachman, A. Trischler, and Y. Bengio.
+430 Learning deep representations by mutual information estimation and maximization. arXiv preprint
+431 arXiv:1808.06670, 2018.
+432 [19] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal
+433 covariate shift. arXiv preprint arXiv:1502.03167, 2015.
+434 [20] H. Kataoka, T. Wakamiya, K. Hara, and Y. Satoh. Would mega-scale datasets further enhance spatiotempo
+435 ral 3d cnns? arXiv preprint arXiv:2004.04968, 2020.
+36 [21] D. Kim, D. Cho, and I. S. Kweon. Self-supervised video representation learning with space-time cubic puzzles. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8545–8552, 2019.
+39 [22] B. Korbar, D. Tran, and L. Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In Advances in Neural Information Processing Systems, pages 7763–7774, 2018. [23] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. Hmdb: a large video database for human motion recognition. In 2011 International Conference on Computer Vision, pages 2556–2563. IEEE, 2011.
+43 [24] Z. Lai and W. Xie. Self-supervised learning for video correspondence flow. arXiv preprint arXiv:1905.00875, 2019.
+5 [25] H.-Y. Lee, J.-B. Huang, M. Singh, and M.-H. Yang. Unsupervised representation learning by sorting sequences. In Proceedings of the IEEE International Conference on Computer Vision, pages 667–676, 2017. [26] D. Mahajan, R. Girshick, V. Ramanathan, K. He, M. Paluri, Y. Li, A. Bharambe, and L. van der Maaten. Exploring the limits of weakly supervised pretraining. In Proceedings of the European Conference on Computer Vision (ECCV), pages 181–196, 2018. [27] A. Miech, J.-B. Alayrac, L. Smaira, I. Laptev, J. Sivic, and A. Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879–9889, 2020. [28] A. Miech, D. Zhukov, J.-B. Alayrac, M. Tapaswi, I. Laptev, and J. Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE international conference on computer vision, pages 2630–2640, 2019. [29] I. Misra and L. v. d. Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707–6717, 2020. [30] I. Misra, C. L. Zitnick, and M. Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European Conference on Computer Vision, pages 527–544. Springer, 2016. [31] P. Morgado, N. Vasconcelos, and I. Misra. Audio-visual instance discrimination with cross-modal agreement. arXiv preprint arXiv:2004.12943, 2020.
+63 [32] V. Nair and G. E. Hinton. Rectified linear units improve restricted boltzmann machines. In ICML, 2010.
+4 [33] M. Noroozi and P. Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European Conference on Computer Vision, pages 69–84. Springer, 2016.
+6 [34] M. Noroozi, H. Pirsiavash, and P. Favaro. Representation learning by learning to count. In Proceedings of the IEEE International Conference on Computer Vision, pages 5898–5906, 2017.
+68 [35] A. v. d. Oord, Y. Li, and O. Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. [36] A. Owens and A. A. Efros. Audio-visual scene analysis with self-supervised multisensory features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631–648, 2018. [37] A. Owens, J. Wu, J. H. McDermott, W. T. Freeman, and A. Torralba. Ambient sound provides supervision for visual learning. In European conference on computer vision, pages 801–816. Springer, 2016. [38] D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, and A. A. Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536–2544, 2016. [39] M. Patrick, Y. M. Asano, R. Fong, J. F. Henriques, G. Zweig, and A. Vedaldi. Multi-modal self-supervision from generalized data transformations. arXiv preprint arXiv:2003.04298, 2020. [40] K. J. Piczak. Esc: Dataset for environmental sound classification. In Proceedings of the 23rd ACM international conference on Multimedia, pages 1015–1018, 2015. [41] S. Purushwalkam and A. Gupta. Demystifying contrastive self-supervised learning: Invariances, augmentations and dataset biases. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 3407–3418, 2020.
+[42] K. Soomro, A. R. Zamir, and M. Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.
+[43] C. Sun, A. Shrivastava, S. Singh, and A. Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843–852, 2017.
+[44] Y. Tian, D. Krishnan, and P. Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849, 2019.
+[45] D. Tran, H. Wang, L. Torresani, J. Ray, Y. LeCun, and M. Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450–6459, 2018.
+[46] C. Vondrick, A. Shrivastava, A. Fathi, S. Guadarrama, and K. Murphy. Tracking emerges by colorizing videos. In Proceedings of the European conference on computer vision (ECCV), pages 391–408, 2018.
+[47] F. Wang and H. Liu. Understanding the behaviour of contrastive loss. arXiv preprint arXiv:2012.09740, 2020.
+[48] J. Wang, J. Jiao, L. Bao, S. He, Y. Liu, and W. Liu. Self-supervised spatio-temporal representation learning for videos by predicting motion and appearance statistics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4006–4015, 2019.
+[49] D. Xu, J. Xiao, Z. Zhao, J. Shao, D. Xie, and Y. Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10334–10343, 2019.
+[50] R. Zhang, P. Isola, and A. A. Efros. Colorful image colorization. In European conference on computer vision, pages 649–666. Springer, 2016.
+# Checklist
+1. For all authors...
+(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]
+(b) Did you describe the limitations of your work? [Yes] Specifically, training data being proprietary creates concerns around reproducibility, which has precedence in the literature as mentioned in the paper. We address that partially by planning to publicly release pretrained models.
+(c) Did you discuss any potential negative societal impacts of your work? [Yes]
+(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
+2. If you are including theoretical results...
+(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A]
+3. If you ran experiments...
+(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No] The data is proprietary. However, we have provided implementation of the proposed method in supplemental material and aim to publicly release the pretrained models.
+(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] They are all discussed in detail either in the main submission or in supplemental material.
+(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] We only observed meaningful differences after running experiments multiple times, for ESC50[40] downstream experiments. Corresponding standard errors are reported in supplemental material.
+(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] Please refer to Section 4.1.
+4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
+(a) If your work uses existing assets, did you cite the creators? [N/A]
+(b) Did you mention the license of the assets? [N/A]
+(c) Did you include any new assets either in the supplemental material or as a URL? [N/A]
+(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A]
+(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A]
+5. If you used crowdsourcing or conducted research with human subjects...
+(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A]
+(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]
+(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A]

parse/train/8hGabvaV2GQ/8hGabvaV2GQ_content_list.json ADDED Viewed

	@@ -0,0 +1,887 @@

+[
+    {
+        "type": "text",
+        "text": "Watching Too Much Television is Good: Self-Supervised Audio-Visual Representation Learning from Movies and TV Shows ",
+        "text_level": 1,
+        "bbox": [
+            227,
+            122,
+            772,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Anonymous Author(s)   \nAffiliation   \nAddress   \nemail ",
+        "bbox": [
+            423,
+            251,
+            578,
+            306
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract ",
+        "text_level": 1,
+        "bbox": [
+            462,
+            343,
+            535,
+            359
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 The abundance and ease of utilizing sound, along with the fact that auditory clues   \n2 reveal so much about what happens in the scene, make the audio-visual space a   \n3 perfectly intuitive choice for self-supervised representation learning. However,   \n4 the current literature suggests that training on uncurated data yields considerably   \n5 poorer representations compared to the curated alternatives collected in supervised   \n6 manner, and the gap only narrows when the volume of data significantly increases.   \n7 Furthermore, the quality of learned representations is known to be heavily influ  \n8 enced by the size and taxonomy of the curated datasets used for self-supervised   \n9 training. This begs the question of whether we are celebrating too early on catching   \n10 up with supervised learning when our self-supervised efforts still rely almost exclu  \n11 sively on curated data. In this paper, we study the efficacy of learning from Movies   \n12 and TV Shows as forms of uncurated data for audio-visual self-supervised learning.   \n13 We demonstrate that a simple model based on contrastive learning, trained on a   \n14 collection of movies and TV shows, not only dramatically outperforms more com  \n15 plex methods which are trained on orders of magnitudes larger uncurated datasets,   \n16 but also performs very competitively with the state-of-the-art that learns from   \n17 large-scale curated data. We identify that audiovisual patterns like the appearance   \n18 of the main character or prominent scenes and mise-en-scène which frequently   \n19 occur through the whole duration of a movie, lead to an overabundance of easy   \n20 negative instances in the contrastive learning formulation. Capitalizing on such   \n21 observation, we propose a hierarchical sampling policy, which despite its simplicity,   \n22 effectively improves the performance, particularly when learning from TV shows   \n23 which naturally face less semantic diversity. ",
+        "bbox": [
+            148,
+            373,
+            766,
+            693
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "24 1 Introduction ",
+        "text_level": 1,
+        "bbox": [
+            148,
+            717,
+            312,
+            734
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "25 Recently, there has been tremendous progress in self-supervised learning from still images, where the   \n26 standard supervised training has been outperformed in a variety of image-related tasks [7, 8, 15, 29].   \n27 The appeal of detaching representation learning from human annotations is rooted not only in the   \n28 non-trivial challenges of scaling-up the labeling process, but also in the ill-defined task of determining   \n29 a proper taxonomy with generalization power and transferability. Both challenges only exacerbate as   \n30 we move from images to videos, where the notion of time is involved and the complexity of visual   \n31 concepts increases. Simply considering the number of training instances or even the cardinality of   \n32 the label set is not sufficient to conclude if one large-scale supervised dataset is more suitable than   \n33 another for transfer learning in video classification tasks [20]. That is, the abundance of attention   \n34 which video self-supervised learning has lately received is only to be expected. While many research   \n35 efforts in this area extend the contributions made initially in the image domain to the video domain,   \n36 others, including our work, have explored harnessing additional modalities such as audio or text for   \n37 multi-modal self-supervised learning [2, 3, 4, 22, 27, 31, 37, 36, 39].   \n38 From the current state-of-the-art one makes two major conclusions. First, the quality of learned   \n39 representations, evaluated by fine-tuning on downstream tasks, is heavily influenced by the size and   \n40 taxonomy of the pretraining datasets [2, 3, 39]. Second, an uncurated pretraining dataset yields   \n41 considerably poorer representations compared to a curated one and the gap only narrows when the   \n42 total amount of pretraining data significantly increases [3]. Curated data refers to likes of supervised   \n43 large-scale action recognition and audio classification datasets such as Kinetics [6], IG-Kinetics   \n44 [12], AudioSet [11], and YouTube-8M [1]. While the human-annotated labels are not accessed for   \n45 self-supervised pretraining, videos being trimmed and from a label set of limited cardinality with   \n46 biased sampling distribution1 implicitly acts as a sort of supervision. On the other hand, an uncurated   \n47 data refers to likes of IG-Random[3], simply a body of unlabeled videos collected blindly with   \n48 none of the aforementioned careful human-involvements. That being said, we know that something   \n49 as simple as having access to a clean object-centric training data, like Imagenet, can be indirectly   \n50 exploited by contrastive self-supervised learning in image domain to obtain additional performance   \n51 gain [41] on the downstream tasks which exhibit similar properties. The analogous to it of course are   \n52 the well trimmed closed-set curated datasets which are being extensively used in the literature for   \n53 video self-supervised pretraining, while downstream evaluations focus on benchmarks with similar   \n54 characteristics. Our work aims at comprehensively exploring the efficacy of learning from Movies   \n55 and TV Shows, as forms of uncurated data, for audio-visual self-supervised learning.   \n56 Many of us can relate to an experience in movie theaters when the sound of the engine, first perceived   \n57 by our left ear, is gradually heard more by the right ear as a car moves from the left side of the   \n58 screen to the right side. Another example is a scene in which an object, like a helicopter, approaches   \n59 the camera from distance and eventually flies over it. In this case, the perceived sound not only   \n60 changes in loudness but also transitions from front to back, in concert with the visuals, giving the   \n61 audience a more realistic feeling as if they are indeed positioned behind the camera. Besides, with   \n62 art being inherently novel, two movies even if they share genres or revolve around similar story   \n63 lines often deliver quite different experiences and portray distinct visuals, thanks to the extremely   \n64 artist-driven creative process behind such productions. We hypothesize that the aforementioned high   \n65 audio fidelity, and inherent semantic diversity characterize long-form content2 as potentially a very   \n66 rich source for self-supervised multi-modal representation learning. It is worth emphasizing that in   \n67 spirit of uncurated data, we not only blindly sample from a large collection of movies and TV shows   \n68 when constructing our pretraining dataset, but also perform ablation studies on the effect of genre   \n69 distribution, the closest we have to taxonomy in the curated datasets, confirming that the quality of   \n70 learned representations is agnostic with respect to such statistics.   \n71 To the best of our knowledge, we are the first to solely rely on uncurated data and study the efficacy of   \n72 self-supervised multi-modal representation learning from movies and TV shows. Despite meaningful   \n73 domain gap between our pretraining data and the space of downstream tasks, we obtain representations   \n74 which are very competitive with those learned from curated datasets. This is particularly important as   \n75 we follow a much simpler modeling approach in comparison with the state-of-the art. ",
+        "bbox": [
+            145,
+            748,
+            825,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            155,
+            92,
+            823,
+            119
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            145,
+            126,
+            825,
+            375
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            145,
+            381,
+            825,
+            587
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            148,
+            594,
+            825,
+            664
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "76 2 Related Work ",
+        "text_level": 1,
+        "bbox": [
+            148,
+            685,
+            320,
+            702
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "77 Self-supervised learning techniques define pretext tasks, mostly inspired by the natural structures   \n78 in the data, in order to generate supervisory signals for training. Despite the plethora of proposed   \n79 pretext tasks in the literature, these approaches can be coarsely divided into two groups, namely   \n80 pretext learning, and pretext-invariant methods. Approaches which fall in the former bucket, usually   \n81 apply a form of transform, randomly drawn from a parametric family, to the input data then optimize   \n82 for predicting the parameters of the chosen transformation. Predicting the relative position of image   \n83 patches [9], solving jigsaw puzzles [33], estimating artificial rotations [13], colorization [50], context   \n84 encoders learned through inpainting [38], and learning by counting scale and split invariant visual   \n85 primitives [34], are among many methods which belong to this category. Similar techniques have   \n86 been extended from images to videos [10, 21, 24, 25, 30, 46, 48, 49], where in addition to the   \n87 spatial context, the temporal domain, and the arrow of time have been heavily exploited. In contrast,   \n88 pretext-invariant methods [5, 7, 8, 15, 18, 17, 29, 35, 39, 44] are built on the concept of maximizing   \n89 mutual information across augmented versions of a single instance, and are mostly formulated as   \n90 contrastive learning. In other words, a pretext is used to generate different views of a single input   \n91 for which the learning algorithm aims to maximize the intra-instance similarity, across variety of   \n92 transformations. Our work falls within this category, however we function in a multi-modal realm   \n93 employing both audio and video.   \n94 Earlier works which harnessed audio and video for representation learning, have leveraged audio  \n95 visual temporal synchronization [22, 36], correspondence [4], and cross-modal clustering [3, 37]. The   \n96 work by Patrick et al.[39] proposes a generalized data transformation in order to unify a variety of   \n97 audio-visual self-supervised pretext tasks through a noise contrastive formulation. This work is close   \n98 to ours in choice of objective function and data type, yet we employ no augmentation (except modality   \n99 projection in the terminology of [39]), and solely focus on capitalizing the advantages of learning   \n100 from long-form content. Morgado et al.[31] show that cross-modal discrimination is important for   \n101 learning good audio and video representations, something which was also pointed out earlier in a   \n102 clustering framework [3]. Beyond that, [31] generalizes the notion of instance-level positive and   \n103 negative examples by exploring cross-modal agreement where multiple instances are grouped together   \n104 as positives by measuring their similarity in both the video and audio feature spaces. While we also   \n105 adopt a cross-modal noise contrastive estimation loss, we stick with the vanilla version, instance-level   \n106 positive and negatives, and do not use any memory bank feature representations. Finally, Alayrac et   \n107 al.[2] recently proposed a multi-modal versatile network capable of simultaneously learning from   \n108 audio, video and text. Building on the intuition that different modalities are of different semantic   \n109 granularity, audio and video are first compared in a fine-grained space while text is compared with   \n110 the aforementioned modalities in a lower dimensional coarse-grained space. In our experiments, we   \n111 compare with a variant of [2] where only audio and video modalities are utilized. ",
+        "bbox": [
+            148,
+            718,
+            825,
+            869
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            145,
+            90,
+            825,
+            174
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            142,
+            181,
+            825,
+            430
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "112 3 Approach ",
+        "text_level": 1,
+        "bbox": [
+            143,
+            449,
+            289,
+            467
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "113 Notations and Architecture. Our pretraining dataset is denoted by $\\mathcal { X } \\ : = \\ : \\{ \\mathcal { X } _ { n } | n \\in [ 1 \\cdot \\cdot \\cdot N ] \\}$   \n114 where $\\mathcal { X } _ { n } = \\{ x _ { n , m } | m \\in [ 1 \\cdots M _ { n } ] \\}$ contains $M _ { n }$ non-overlapping audiovisual snippets which are   \n115 temporally segmented from the duration of the $n ^ { t h }$ long-form content in the dataset. Each snippet   \n116 includes both audio and video modalities, formally $x _ { n , m } = ( a _ { n , m } , v _ { n , m } )$ , where $a _ { n , m } \\in \\mathbb { R } ^ { 1 \\times \\pmb { \\tilde { P } } \\times Q }$   \n117 and $v _ { n , m } \\in \\mathbb { R } ^ { 3 \\times T \\times H \\times W }$ . $T , H$ , and $W$ denote the number of frames, height and width of the video,   \n118 while $P$ , and $Q$ respectively stand for the number of mel filters, and audio frames. Video and audio are   \n119 processed through 18-layers deep $\\mathrm { R } ( 2 { + } 1 ) \\mathrm { D }$ [45] and ResNet [16] architectures, respectively referred   \n120 to as $f : \\mathbb { R } ^ { 3 }  \\check { \\mathbb { R } } ^ { d _ { f } }$ and $\\dot { \\boldsymbol { g } } : \\mathbb { R } ^ { 1 } \\dot {  } \\mathbb { R } ^ { d _ { g } }$ . Inspired by [7], we use projection heads, $\\boldsymbol { h } _ { f } : \\mathbb { R } ^ { d _ { f } } \\overset { \\cdot } {  } \\mathbb { R } ^ { d }$ and   \n121 $h _ { g } : \\mathbb { R } ^ { d _ { g } }  \\mathbb { R } ^ { d }$ , to map corresponding representations into a common $d$ -dimensional space before   \n122 computing the contrastive loss. The shallow architecture of $h _ { f }$ and $h _ { g }$ consists of two convolution   \n123 layers, separated by Batch Normalization [19] and ReLU [32], followed by global average pooling.   \n124 Once self-supervised pretraining finished, we discard the projection heads and fine-tune $f$ and $g$ for   \n125 respective downstream tasks.   \n126 Loss Function. With a slight abuse of notation3, $\\boldsymbol { B } = \\{ x _ { i } = ( a _ { i } , v _ { i } ) | i \\in [ 1 \\cdot \\cdot \\cdot B ] \\}$ represents a   \n127 minibatch of size $B$ , where video and audio modalities associated with the $i ^ { t h }$ sample, $x _ { i }$ , are denoted   \n128 by $v _ { i }$ and $a _ { i }$ . We use $z _ { v } ^ { i } = h _ { f } ( f ( v _ { i } ) )$ and $z _ { a } ^ { i } = h _ { g } ( g ( a _ { i } ) )$ to represent the associated embeddings   \n129 generated by projection heads, and optimize the noise-contrastive loss [14] shown in 1 in order   \n130 to maximize the symmetric joint probability between audio and video. For the $i ^ { t h }$ element in the   \n131 minibatch, $( z _ { v } ^ { i } , z _ { a } ^ { i } )$ serves as the positive pair, while assuming negative pairs for both modalities,   \n132 $\\mathcal { N } _ { i } = \\{ ( z _ { v } ^ { i } , z _ { a } ^ { j } ) , ( \\stackrel { \\sim } { z } _ { v } ^ { j } , z _ { a } ^ { i } ) | j \\in [ 1 \\cdots B ] , i \\neq j \\}$ constitutes the set of negative pairs. ",
+        "bbox": [
+            140,
+            481,
+            825,
+            667
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            140,
+            671,
+            826,
+            773
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/8fcb2807e4a2e1c580b4b3eec9c90510b9e6e966a2703c3454b59363ad66f489.jpg",
+        "text": "$$\n\\mathcal { L } = - \\sum _ { i = 1 } ^ { B } \\log \\left( \\frac { e ^ { ( z _ { v } ^ { i } ) ^ { \\top } ( z _ { a } ^ { i } ) } } { e ^ { ( z _ { v } ^ { i } ) ^ { \\top } ( z _ { a } ^ { i } ) } + \\displaystyle \\sum _ { ( z _ { v } ^ { \\prime } , z _ { a } ^ { \\prime } ) \\in \\mathcal { N } _ { i } } e ^ { ( z _ { v } ^ { \\prime } ) ^ { \\top } ( z _ { a } ^ { \\prime } ) } } \\right)\n$$",
+        "text_format": "latex",
+        "bbox": [
+            325,
+            787,
+            671,
+            847
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "133 Most of the previous works [2, 31, 39] normalize the embeddings before computing the contrastive   \n134 loss and employ a temperature hyper-parameter, often denoted by $\\tau$ as in [2, 31], to control the   \n35 smoothness for the distribution of pairwise similarities. In contrast, we have chosen to operate in an   \n36 unnormalized embedding space. Besides the obvious benefit of eliminating the need for tuning $\\tau$ , we   \n7 empirically show that such decision does not affect the quality of the learned representations.   \n138 Sampling Policy. Contrastive loss function shown in Equation 1 is computed over $B$ training   \n139 instances, each in form of an audiovisual snippet. A naive sampling policy may ignore the fact that   \n140 snippets comprising the pretraining dataset are in fact temporal segments that were trimmed from   \n141 longer-form contents, i.e. movies and TV shows. Such an assumption treats our training data as   \n142 independent and identically distributed random variables from $\\textstyle \\bigcup _ { n = 1 } ^ { N } { \\mathcal { X } } _ { n }$ , which constitutes the default   \n143 sampling policy that is commonly used in the general deep learning literature. However, in reality,   \n144 commonalities and correlations do exist along the temporal axis of a movie or TV show, things like   \n145 audio mastering artifacts, frequent appearance of the main character’s face and voice, thematic music,   \n146 repetitive scenes and mise-en-scène4, all of which contribute to breaking the previously discussed   \n147 i.i.d assumption. This is even more pronounced when we deal with multiple episodes of the same TV   \n148 show appearing in the pretraining dataset5. Note that, sampling from no video data is going to be i.i.d   \n149 but in this case the temporal correlations extend for much longer given our entities are movies and TV   \n150 shows. Thus, it is more accurate to think of $\\mathcal { X }$ having multiple underlying domains, oriented towards   \n151 exclusive properties which different long-form contents are characterized by. We hypothesize that   \n152 during training, model gradually discovers such patterns of commonalities, which are not semantically   \n153 valuable, and latches onto those to quickly minimize Equation 1 leading to poor generalization6. The   \n154 reason being $B \\ll N$ , hence for $n \\sim \\mathbb { U } ( 1 , N )$ and $m \\neq m ^ { \\prime }$ , $\\mathsf { P } ( x _ { n , m } \\in B \\land x _ { n , m ^ { \\prime } } \\in B )$ is negligible.   \n155 In other words, the set of negative pairs in Equation 1 mainly includes pairs for which audio and   \n156 video come from two different movies or TV shows, thus due to the aforementioned artifacts behave   \n157 as easy negatives.   \n158 In order to quantitatively measure our hypothesis, we define different distributions, shown in Equation   \n159 2, over the space of audio-visual similarity. $S ^ { + }$ indicates the space of correct matches, i.e. where   \n160 audio and video correspond to the same snippet. $S ^ { - }$ indicates the space where audio and video do   \n161 not correspond yet belong to the same movie or TV show. Finally, $\\bar { \\boldsymbol { S } } ^ { \\neq }$ indicates the space in which   \n162 audio and video are sampled from two distinct long-form content, hence naturally do not correspond. ",
+        "bbox": [
+            138,
+            857,
+            825,
+            886
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            153,
+            90,
+            823,
+            133
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            138,
+            137,
+            826,
+            417
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            140,
+            424,
+            825,
+            494
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "img_path": "images/6c6c65f24d0869d565b508cfc76b159ac2970f28edea5117d9f0294494218988.jpg",
+        "text": "$$\n( z _ { v } ^ { n , m } ) ^ { \\mathsf { T } } ( z _ { a } ^ { n ^ { \\prime } , m ^ { \\prime } } ) \\sim \\left\\{ { \\begin{array} { l l } { S ^ { + } , } & { { \\mathrm { i f } } \\ n = n ^ { \\prime } \\wedge m = m ^ { \\prime } } \\\\ { S ^ { - } , } & { { \\mathrm { i f } } \\ n = n ^ { \\prime } \\wedge m \\neq m ^ { \\prime } } \\\\ { S ^ { \\neq } , } & { { \\mathrm { i f } } \\ n \\neq n ^ { \\prime } \\wedge \\forall ( m , m ^ { \\prime } ) } \\end{array} } \\right.\n$$",
+        "text_format": "latex",
+        "bbox": [
+            331,
+            511,
+            665,
+            563
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "163 With that, and $\\mathsf { K L }$ denoting Kullback–Leibler divergence, $\\mathsf { K L } ( S ^ { - } \\parallel S ^ { + } )$ measures the expected   \n164 difference between positive and negative pairs within the same movie or TV show. Ideally, this should   \n165 increase as the training progresses, since the model gradually learns audio-video correspondence by   \n166 minimizing Equation 1. Meanwhile, the i.i.d assumption suggests $\\mathsf { K L } ( \\mathcal { S } ^ { - } \\parallel \\mathcal { S } ^ { + } ) \\simeq \\mathsf { K L } ( \\mathcal { S } ^ { \\neq } \\parallel \\mathcal { S } ^ { + } )$   \n167 and $\\mathsf { K L } ( \\mathcal { S } ^ { - } \\parallel \\mathcal { S } ^ { \\neq } ) \\simeq \\boldsymbol { 0 }$ , yet as we empirically illustrate later, $\\mathsf { K L } ( \\mathcal { S } ^ { - } \\parallel \\mathcal { S } ^ { + } ) < \\mathsf { K L } ( \\mathcal { S } ^ { \\neq } \\parallel \\mathcal { S } ^ { + } )$ and   \n168 ${ \\mathsf { K L } } ( S ^ { - } \\parallel S ^ { \\neq } )$ is rather large, indicating that, upon convergence and on a held-out set, model has   \n169 a harder time pushing apart negative pairs when audio and video come from the same underlying   \n170 long-form content. Next, we explain how a simple alternative policy which samples $k$ snippets   \n171 from each long-form content effectively reduces both of the discrepancy measures, referring to   \n172 $\\mathsf { K L } ( S ^ { - } \\parallel S ^ { \\neq } )$ and $\\mathsf { K L } ( \\mathcal { S } ^ { \\neq } \\parallel \\mathcal { S } ^ { + } ) - \\mathsf { K L } ( \\mathcal { S } ^ { - } \\parallel \\mathcal { S } ^ { + } )$ , while yielding better generalization on a range   \n173 of downstream tasks.   \n174 To ameliorate the aforementioned optimization challenge, we take a hierarchical approach. In   \n175 particular, we first uniformly sample a long-form content, $n \\sim \\mathbb { U } ( 1 , N )$ , and then draw $k$ distinct   \n176 snippets from ${ \\mathcal { X } } _ { n }$ , creating $\\{ { \\bar { x } } _ { n , m } | { \\bar { m } } \\in { \\mathcal { M } } _ { n } \\}$ , where $\\mathcal { M } _ { n } \\subset [ 1 \\cdots M _ { n } ]$ and $| { \\mathcal { M } } _ { n } | = k$ . This ensures   \n177 that for $x _ { i } \\in B$ , ${ \\mathcal { N } } _ { i }$ always includes $2 k - 2$ pairs sampled from the same movie or TV show to   \n178 which $x _ { i }$ belongs. By putting constraints on $\\mathcal { M } _ { n }$ , specifically how temporally far from each other   \n179 the $k$ samples are drawn, we may go one step further and to some extent control the audiovisual   \n180 similarity between snippets. This serves as an additional nob to tune for hard negative sampling.   \n181 The intuition is that, the larger narrative of a professionally made movie or TV show is composed of   \n182 shorter units called scene. Each scene comprises a complete event, action, or block of storytelling and   \n183 normally takes place in one location and deals with one action. That is, if our samples are temporally   \n184 close, it is more likely for corresponding snippets to be highly correlated and/or look/sound alike.   \n185 $k \\leq \\mathrm { m a x } [ \\mathcal { M } _ { n } ] - \\mathrm { m i n } [ \\mathcal { M } _ { n } ] + 1 \\ \\bar { \\leq } \\ w \\leq \\bar { M } _ { n }$ defines the bounds on our sampling policy, where $w$   \n186 standing for a sampling window, determines the farthest two out of $k$ samples drawn from ${ \\mathcal { X } } _ { n }$ can   \n187 be. Accordingly, $w = k$ represents the case where all $k$ samples are temporally adjacent, hence   \n188 the expected audiovisual similarity is maximized due to temporal continuity in content. We show   \n189 that having such level of hard negatives, even with a small $k$ , prevents proper training and results in   \n190 performance degradation. On the other hand, $w = M _ { n }$ indicates random sampling where no temporal   \n191 constraint is imposed on $\\mathcal { M } _ { n }$ , thus samples are less likely to be drawn from adjacent time-stamps.   \n192 In this case, expected audiovisual similarity (i.e. hardness of negative pairs) is mainly derived from   \n193 global content-exclusive artifacts like, color palette, frequent appearance of the main character’s face   \n194 and voice, repetitive scenes, and etc. The rest of the spectrum provides middle grounds where two   \n195 samples drawn from ${ \\mathcal { X } } _ { n }$ can at most be $w + 1$ snippets apart, something reminiscent of temporal   \n196 locality. Our sampling policy can be easily implemented in a few lines of Python. Please refer to   \n197 supplemental material for further details. ",
+        "bbox": [
+            140,
+            573,
+            825,
+            729
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            145,
+            734,
+            825,
+            861
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            140,
+            92,
+            825,
+            297
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "98 4 Experiments ",
+        "text_level": 1,
+        "bbox": [
+            156,
+            318,
+            312,
+            334
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experimental Setup ",
+        "text_level": 1,
+        "bbox": [
+            173,
+            348,
+            351,
+            363
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets and Reproducibility. We use full-length movies and episodes of TV shows for selfsupervised pretraining. Titles are randomly chosen from a large collection spanning over a variety of genres, namely Drama, Comedy, Action, Horror, Thriller, Sci-Fi and Romance. All audio is in English language. Our Movie dataset, consists of 3.6K films with an average duration of 105 minutes. Our TV dataset includes 9.2K episodes from a total of 581 shows with an average duration of 42 minutes per episode. Each of our datasets comprises 0.7 years worth of uncurated audiovisual content, which is significantly smaller than IG-Random [3] with variants at 5 and 21 years. Scaling up our pretraining datasets to volumes comparable to the IG-Random [3] while possible is non-trivial and demands dramatically larger compute resources for training, something which we currently cannot afford. Given that we cannot publicly release our dataset due to copyright reasons, we acknowledge that it is not possible for other research groups to fully reproduce our results. However, we intend to make available the pretrained models and hope that research community finds them, along with the other contributions of this work, of value whether within the context of self-supervised learning or adoption for various downstream tasks. We would like to emphasize that similar limitations have precedents in multiple earlier works including but not limited to [3, 12, 26, 43]. To evaluate the efficacy of self-supervised audio-visual representation learning from movies and TV shows, we follow recent works [3, 39, 31, 2] and benchmark UCF101[42] and HMDB51[23] for action recognition, along with ESC50[40] for audio classification. Results for the ablation studies are reported on the split-1 of the corresponding datasets. Following the standard protocol, we report the average performance over all splits when we are comparing with the state-of-the-art. ",
+        "bbox": [
+            171,
+            372,
+            825,
+            650
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Pretraining. Unless mentioned otherwise, we use video snippets with 16 frames at 5 fps. For data augmentation, we resize the shorter side to 190 pixels, then randomly crop them into $1 5 8 \\times 1 5 8$ pixels. As for sound, we compute mel spectrogram from the raw audio at 48K sample rate using $9 6 ~ \\mathrm { m e l }$ filters and an FFT window of 2048, while the number of samples between successive frames is set to 512. For data augmentation, we randomly drop out up to $2 5 \\%$ from either temporal or frequency axis of the 2-D mel spectrogram image. Training uses a batch size of 512 and takes on average 42 hours on 8 NVIDIA A100 GPUs. The dimension of audio-video joint embedding space, $d$ , is set to 512. ",
+        "bbox": [
+            166,
+            656,
+            825,
+            752
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Downstream Evaluation. For training on UCF101 [42] and HMDB51 [23], we use video clips that are 32 frames long at 10 fps. Unless mentioned otherwise, these clips are randomly chosen from the duration of the video instances. A scale jittering range of [181, 226] pixels is used and we randomly crop the video into $1 5 8 \\times 1 5 8$ pixels. Furthermore, random horizontal flipping and color jittering are employed. During inference, 10 temporal clips are uniformly sampled where each is spatially cropped in 3 ways (left, center, right) resulting in a total of 30 views. We then average the model predictions across these 30 views and report top-1 classification accuracy. For training on ESC50 [40], we use 3-seconds clips which are randomly chosen from the duration of the audio instances and apply time and frequency masking to spectrogram images for data augmentation. The maximum possible length of the mask is $50 \\%$ of the corresponding axis. We do not use any scale jittering or random cropping on the spectrograms. During inference, 10 temporal clips are uniformly sampled ",
+        "bbox": [
+            171,
+            758,
+            825,
+            911
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "238 and we average the model predictions across these 10 views and report top-1 classification accuracy.   \n239 For further implementation details, please refer to the supplemental material. ",
+        "bbox": [
+            143,
+            90,
+            826,
+            119
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/fe3fa73a08dcc0cf4d8abc90b46bc44d9e947f10dd7a4be6b2c5ef8be9fc6fa5.jpg",
+        "image_caption": [
+            "Figure 1: Ablation study of the proposed sampling policy on reducing the discrepancy measures. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            233,
+            136,
+            764,
+            531
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "240 4.2 Ablation Study ",
+        "text_level": 1,
+        "bbox": [
+            142,
+            579,
+            318,
+            594
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In the following, we discuss multiple ablation studies to assess our main hypothesis that, a hierarchical sampling policy, as described in Section 3, enables better representations to be learned by increasing the portion of hard negative pairs which the contrastive loss function observes. Here, pretraining uses $90 \\%$ of either Movie or TV dataset, while the remaining $10 \\%$ constitute a held-out validation set7 on which we report the discrepancy measures. ",
+        "bbox": [
+            173,
+            604,
+            547,
+            715
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Sample size $( k )$ Figure 1a illustrates that compared to the baseline sampling denoted by $k = 1$ , our approach $k > 1 ,$ ) effectively shrinks the gap between $S ^ { - }$ and ${ \\mathcal { S } } ^ { \\neq }$ when measured either directly or against $S ^ { + }$ . Its pattern of behavior also perfectly follows our earlier intuition (ref. Section 3). In particular, given a fixed minibatch budget, a larger $k$ favors more training instances to be sampled from fewer number of long-form contents. That increases the portion of hard negative pairs, thus pushes the contrastive loss to more aggressively separate mismatched audio-video pairs from the same movie, which leads model to maintain less of the content-exclusive artifacts in the embedding space. In the most extreme case, $k = 6 4$ , all the training instances are sampled from the same movie. From Table 1, we observe that different variants of our sampling policy, with no imposed temporal constraint, i.e. $w = M _ { n }$ , outperform the baseline on all three downstream tasks ",
+        "bbox": [
+            173,
+            722,
+            547,
+            859
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/bf6588579202158118301afb6941dec79d9e5cdccd64dae9b2636f3ddcd67c90.jpg",
+        "image_caption": [
+            "Figure 2: Effect of color jitter on the discrepancy measures. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            618,
+            821,
+            801
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            178,
+            859,
+            825,
+            888
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            163,
+            90,
+            823,
+            133
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Sampling window $\\mathbf { \\Pi } ^ { ( w ) }$ Smaller $w$ forces samples that belong to same movie to be drawn from a shorter temporal window, hence growing the probability that they look/sound very much alike (i.e. harder negative pairs). That is, it should further diminish the discrepancy measures. Figure 1b illustrates this behavior where we gradually increase $w$ while $k = 1 6$ . However, from Table 1, it does not seem that tuning for $w$ , i.e $w \\neq M _ { n }$ , provides a meaningful gain on downstream tasks. This implies that commonalities which persist throughout the duration of a movie are sufficiently powerful signals to be exploited for generating hard negatives. We hypothesize that different scenes both within and across different movies and TV shows are of variety of length, thus a fixed $w$ is sub-optimal. Ideally, we should identify scene boundaries and and dynamically modify $w$ during sampling, something which we leave for future iterations of this work. ",
+        "bbox": [
+            173,
+            138,
+            511,
+            401
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Temporally adjacent samples. Along the lines of previous observations, Figure 1c shows that indeed drawing temporally adjacent snippets from the same long-form content, i.e. $w = k$ , results in aggressively reducing the discrepancy measures. This behavior is agnostic with respect to $k$ yet exacerbates as $k$ grows. Note that, the contrastive loss is an instance-discrimination objective function. Therefore, forcing it to distinguish between temporally adjacent snippets, that naturally sound and look extremely similar, leaves no choice for the model but to discard valuable semantic notions, which predictably leads to poor representations, also confirmed by result reported in Table 1. ",
+        "bbox": [
+            174,
+            409,
+            511,
+            602
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/8666f96fffa6d92d838f89e334ae23434fdf273fe4cc630a04b82155ad5c62a9.jpg",
+        "table_caption": [
+            "Table 1: Ablation study of the proposed sampling policy on different downstream tasks, measured by top-1 classification accuracy. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"5\">pretraining dataset: Movie</td></tr><tr><td>k</td><td>w</td><td>HMDB51</td><td>ESC50</td><td>UCF101</td></tr><tr><td>1</td><td>1</td><td>60.32</td><td>86.50</td><td>85.69</td></tr><tr><td>4</td><td>Mn</td><td>61.37</td><td>89.91</td><td>85.38</td></tr><tr><td>8</td><td>Mn</td><td>62.09</td><td>88.75</td><td>86.06</td></tr><tr><td>16</td><td>Mn</td><td>62.92</td><td>88.33</td><td>86.30</td></tr><tr><td>32</td><td>Mn</td><td>61.04</td><td>88.00</td><td>85.98</td></tr><tr><td>64</td><td>Mn</td><td>61.30</td><td>86.83</td><td>85.43</td></tr><tr><td>16</td><td>64</td><td>60.26</td><td>87.00</td><td>83.61</td></tr><tr><td>16</td><td>128</td><td>60.58</td><td>86.50</td><td>85.30</td></tr><tr><td>16</td><td>256</td><td>62.02</td><td>87.75</td><td>84.85</td></tr><tr><td>16</td><td>512</td><td>61.30</td><td>87.08</td><td>85.38</td></tr><tr><td>16</td><td>1024</td><td>60.65</td><td>86.16</td><td>84.61</td></tr><tr><td>16</td><td>2048</td><td>61.83</td><td>87.66</td><td>85.11</td></tr><tr><td>4</td><td>4</td><td>60.19</td><td>88.00</td><td>84.66</td></tr><tr><td>16</td><td>16</td><td>56.86</td><td>88.75</td><td>82.71</td></tr><tr><td>64</td><td>64</td><td>57.45</td><td>84.58</td><td>82.68</td></tr><tr><td colspan=\"5\">pretraining dataset:TV</td></tr><tr><td>k</td><td>w</td><td>HMDB51</td><td>ESC50</td><td>UCF101</td></tr><tr><td>1</td><td>1</td><td>56.40</td><td>85.50</td><td>84.37</td></tr><tr><td>8</td><td>Mn</td><td>61.50</td><td>87.50</td><td>85.96</td></tr><tr><td>16</td><td>Mn</td><td>61.69</td><td>89.00</td><td>85.64</td></tr><tr><td>8</td><td>64</td><td>60.58</td><td>88.00</td><td>85.96</td></tr><tr><td>8</td><td>128</td><td>60.00</td><td>85.66</td><td>85.77</td></tr><tr><td>16</td><td>256</td><td>61.30</td><td>86.41</td><td>85.01</td></tr></table>",
+        "bbox": [
+            524,
+            204,
+            821,
+            582
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Movies vs. TV Shows. To confirm that our sampling policy behaves consistently across both movies and TV shows, Figure 1d illustrates the discrepancy measures computed on TV dataset. We observe similar effectiveness when using $k$ and $w$ as tuning nobs for reducing either ${ \\mathsf { K L } } ( S ^ { - } \\parallel S ^ { \\neq } )$ or the gap between ${ \\mathsf { K L } } ( S ^ { - } \\parallel S ^ { + } )$ and $\\mathsf { K L } ( S ^ { \\neq } \\parallel S ^ { + } )$ . Table 1 demonstrates that different variants of our approach significantly outperform the baseline, i.e. $k = 1$ . We attribute the larger gains achieved when using TV instead of Movie dataset to the fact that content diversity is naturally lower when pretraining on TV shows since each one includes many episodes that all are characterized with the same content-exclusive artifacts. ",
+        "bbox": [
+            173,
+            608,
+            825,
+            719
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Color jitter. We have established so far that commonalities which persist throughout the duration of a long-form content, things likely associated with color pallet, frequent appearance of the main character’s face and voice, and repetitive scenes can be exploited for learning better representations. That is, one may naturally assume that employing data augmentation techniques like color jitter should be helpful since by distorting content-exclusive visual artifacts, color jitter is expected to reduce ${ \\mathsf { K L } } ( S ^ { - } \\parallel S ^ { \\neq } )$ . Figure 2 illustrates the effect of color jitter, where brightness, contrast, and saturation jitter values are chosen uniformly from $\\mathtt { [ m a x ( 0 , 1 - \\sigma ) , 1 + \\sigma ] }$ . We observe that color jitter reduces the discrepancy measures for the baseline but not as much as it can be obtained by our proposed sampling policy $( k > 1 )$ ), and even then according to Table 2 only yields a slight gain on downstream tasks. ",
+        "bbox": [
+            171,
+            726,
+            825,
+            864
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "$\\ell _ { 2 }$ -normalized feature space. The common practice [2, 31, 39, 7] is to compute contrastive loss in $\\ell _ { 2 }$ -normalized feature space, where according to [47] the temperature hyper-parameter, $\\tau$ , controls the strength of penalties on hard negative samples. We explored this with two widely-used $\\tau$ values. From Table 2, we observe that compared to operating in an unnormalized embedding space, adopting such design choice results in a large performance drop on HMDB51[23] while other downstream benchmarks see only negligible gains. ",
+        "bbox": [
+            147,
+            871,
+            826,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/967321c61bd847c2ad769a0d6a8e812d022c024566f25112b57f3e2a663d6b77.jpg",
+        "table_caption": [
+            "Table 4: Effect of self-supervised learning from curated versus uncurated data on different downstream tasks. The “years” column indicates the duration of the pretraining datasets in years. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>method</td><td>pretraining dataset</td><td>uncurated</td><td>years</td><td>HMDB51</td><td>ESC50</td><td>UCF101</td></tr><tr><td>Ours</td><td>Movie</td><td></td><td>0.7</td><td>62.9</td><td>88.3</td><td>86.3</td></tr><tr><td>Ours</td><td>TV</td><td></td><td>0.7</td><td>61.7</td><td>89.0</td><td>85.6</td></tr><tr><td>XDC[3]</td><td>IG-Random16M</td><td>√</td><td>5</td><td>55.2</td><td>84.3</td><td>84.1</td></tr><tr><td>XDC[3]</td><td>IG-Random65M</td><td>√</td><td>21</td><td>61.2</td><td>86.3</td><td>88.8</td></tr><tr><td>XDC[3]</td><td>IG-Kinetics16M</td><td>×</td><td>5</td><td>57.3</td><td>82.5</td><td>87.6</td></tr><tr><td>XDC[3]</td><td>IG-Kinetics65M</td><td>×</td><td>21</td><td>63.1</td><td>84.8</td><td>91.5</td></tr></table>",
+        "bbox": [
+            230,
+            125,
+            767,
+            237
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            142,
+            262,
+            825,
+            319
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Curated vs. Uncurated data. To the best of our knowledge, the only other uncurated dataset used for audio-visual self-supervised learning is IGRandom[3]8. Table 4 confirms that learning from uncurated movies and TV shows is extremely effective. Our results significantly exceed those of XDC[3] obtained on IG-Random16M despite using a simpler model and 7 times smaller volume of pretraining data. Even in comparison to IG-Random65M with 30 times larger data, we obtain better performances on 2 out of 3 benchmarks. The most promising of our findings though is how competitive our results are against XDC[3] when it is trained on variants of IG-Kinetics which are not only curated but also orders of magnitude larger. With all that, we confidently reject the notion that audio-visual self-supervised learning from uncurated data considerably lags behind utilizing large-scale curated datasets. ",
+        "bbox": [
+            165,
+            325,
+            511,
+            587
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of genre. The distribution of genre among movies used in our pretraining is the closest we have to taxonomy in the curated datasets. So, it is worth examining the quality of our learned representations under various genre distributions. To do so, given a fixed pretraining budget $( N = 1 . 6 \\mathsf { K } )$ , we compare four different scenarios where movies used in the pretraining are distributed i) non-uniformly over all genres except Drama, and Comedies, ii) non-uniformly over Drama, and Comedies, iii) uniformly over all genres, and iv) non-uniformly over all genres. Table 3 confirms that indeed there is very little difference between the aforementioned setups when it comes to transfer learning to the downstream tasks. ",
+        "bbox": [
+            174,
+            593,
+            509,
+            744
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/00e947674b7b31b2c2c93b7df167a033c67701e8a527f8b15b8b88a2bafec0c1.jpg",
+        "table_caption": [
+            "Table 2: Effect of color jitter $( \\sigma )$ and computing contrastive loss in $\\ell _ { 2 }$ -normalized embedding space with temperature hyper-parameter $( \\tau )$ on different downstream tasks. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>k</td><td>0</td><td>HMDB51</td><td>ESC50</td><td>UCF101</td></tr><tr><td>1</td><td>0.0</td><td>60.32</td><td>86.50</td><td>85.69</td></tr><tr><td>16</td><td>0.0</td><td>62.92</td><td>88.33</td><td>86.30</td></tr><tr><td>1</td><td>1.0</td><td>60.45</td><td>87.66</td><td>84.82</td></tr><tr><td>16</td><td>0.5</td><td>60.13</td><td>87.75</td><td>85.98</td></tr><tr><td>16</td><td>1.0</td><td>61.11</td><td>88.33</td><td>85.93</td></tr><tr><td>k</td><td>T</td><td>HMDB51</td><td>ESC50</td><td>UCF101</td></tr><tr><td>16</td><td>0.07</td><td>60.78</td><td>87.08</td><td>86.86</td></tr><tr><td>16</td><td>0.30</td><td>60.78</td><td>89.25</td><td>85.72</td></tr></table>",
+        "bbox": [
+            524,
+            405,
+            818,
+            553
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/200b8c5e62772fbb13faf83321666d8470b1a594fc506d69d09c86a2bd0a944e.jpg",
+        "table_caption": [
+            "Table 3: Effect of genre distribution in Movie dataset on different downstream tasks. Experiments are conducted with input spatial resolution of $1 1 2 \\times 1 1 2$ pixels. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>setting</td><td>HMDB51</td><td>ESC50</td><td>UCF101</td></tr><tr><td>i</td><td>57.58</td><td>86.50</td><td>82.44</td></tr><tr><td>ii</td><td>56.99</td><td>85.50</td><td>82.39</td></tr><tr><td>i</td><td>56.27</td><td>85.25</td><td>82.87</td></tr><tr><td>iv</td><td>56.40</td><td>86.75</td><td>83.24</td></tr></table>",
+        "bbox": [
+            535,
+            631,
+            810,
+            718
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            746,
+            825,
+            773
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.3 Comparison with state-of-the-art ",
+        "text_level": 1,
+        "bbox": [
+            171,
+            790,
+            442,
+            805
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5 compares our proposed approach of learning from Movies and TV shows against the best performing audio-visual self-supervised learning methods. In general, our numbers are comparable with the best existing results reported in the literature, even with much less data and considerably simpler model/training procedure9. It is interesting that training on Movie dataset alone obtains ",
+        "bbox": [
+            174,
+            815,
+            825,
+            872
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/fc2b8e6271d8772cc6c0a709681355292ed496386c4d54260c4cc165203770ed.jpg",
+        "table_caption": [
+            "Table 5: Comparison with state-of-the-art. Dataset abbreviations: AudioSet[11], HowTo100M[28], IG-Kinetics65M [12]; their length in years is given in the “years” column. “Arch.” denotes the architecture of video backbone $( f )$ . [2]† indicates when the corresponding model use only audio and video, and not text modality. For a fair comparison, when using only Movie dataset, we train for twice as many epochs as our other variants in order to match their total number of gradient updates. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Arch.</td><td>pretraining dataset</td><td>curated</td><td>years</td><td>HMDB51</td><td>UCF101</td><td>ESC50</td></tr><tr><td>GDT[39]</td><td>R(2+1)D-18</td><td>AS</td><td></td><td>1</td><td>66.1</td><td>92.5</td><td>88.5</td></tr><tr><td>GDT[39]</td><td>R(2+1)D-18</td><td>IG65M</td><td></td><td>21</td><td>72.8</td><td>95.2</td><td></td></tr><tr><td>XDC[3]</td><td>R(2+1)D-18</td><td>AS</td><td>√</td><td>1</td><td>61.0</td><td>91.2</td><td>84.8</td></tr><tr><td>XDC[3]</td><td>R(2+1)D-18</td><td>IG65M</td><td>√</td><td>21</td><td>67.4</td><td>94.2</td><td></td></tr><tr><td>AVTS[22]</td><td>MC3</td><td>AS</td><td></td><td>1</td><td>61.6</td><td>89.0</td><td>82.3</td></tr><tr><td>AVID[31]</td><td>R(2+1)D-18</td><td>AS</td><td></td><td>1</td><td>64.7</td><td>91.5</td><td>89.1</td></tr><tr><td>MMV[2]+</td><td>R(2+1)D-18</td><td>AS</td><td>√</td><td>1</td><td>70.1</td><td>91.5</td><td>85.6</td></tr><tr><td>MMV[2]+</td><td>S3D-G</td><td>AS</td><td>√</td><td>1</td><td>68.2</td><td>90.1</td><td>86.1</td></tr><tr><td>MMV[2]+</td><td>S3D-G</td><td>AS+HT</td><td>√</td><td>16</td><td>68.3</td><td>91.1</td><td>87.2</td></tr><tr><td>Ours (k=16)</td><td>R(2+1)D-18</td><td>Movie</td><td>X</td><td>0.7</td><td>64.5</td><td>87.9</td><td>88.8</td></tr><tr><td>Ours (k=8)</td><td>R(2+1)D-18</td><td>Movie+TV</td><td>X</td><td>1.4</td><td>65.0</td><td>87.7</td><td>89.1</td></tr><tr><td>Ours (k=16)</td><td>R(2+1)D-18</td><td>Movie+TV</td><td>X</td><td>1.4</td><td>65.1</td><td>88.5</td><td>89.1</td></tr><tr><td>Ours (k=32)</td><td>R(2+1)D-18</td><td>Movie+TV</td><td>X</td><td>1.4</td><td>65.6</td><td>88.7</td><td>88.2</td></tr></table>",
+        "bbox": [
+            176,
+            166,
+            820,
+            369
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "358 comparable performance to the cases where both TV and Movie datasets are used for pretraining.   \n359 This further confirms the richness of the training data which movies and TV shows can provide   \n360 to self-supervised learning problems. We also see that increasing $k$ even beyond 8 gives further   \n361 incremental gains on action recognition benchmarks. ",
+        "bbox": [
+            140,
+            401,
+            825,
+            458
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "362 5 Conclusion ",
+        "text_level": 1,
+        "bbox": [
+            145,
+            486,
+            299,
+            503
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "363   \n364   \n365   \n366   \n367   \n368   \n369   \n370   \n371   \n372   \n373 ",
+        "bbox": [
+            140,
+            522,
+            163,
+            676
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Despite its amazing recent progress, state-of-the-art self-supervised learning still heavily relies on supervised, i.e. curated, large-scale datasets for pretraining. In this work, we have shown that pretraining solely on uncurated data in forms of movies and TV shows, even at a comparatively small scale, can give rise to representations which are capable of competing with the state-of-theart of more complex architectures trained on larger curated datasets. This comes contrary to the current literature which tends to suggest that learning from uncurated data largely falls behind the use of curated alternatives. We intentionally made design decisions to keep our approach and training strategy as simple as possible to demonstrate that learning decently powerful audio-visual representations does not necessarily require gigantic data and compute resources. Through extensive set of experiments, our work establishes for the first time the efficacy of self-supervised learning of audio-visual representations from movies and TV shows. ",
+        "bbox": [
+            148,
+            522,
+            825,
+            674
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "374 6 Broader impact ",
+        "text_level": 1,
+        "bbox": [
+            143,
+            702,
+            336,
+            719
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Potential benefits. Our work shows that competitive multimodal representations can be learned from a comparatively small volume of uncurated data in the form of movies and TV shows. Besides minimizing any sort of human-involvement, which we believe must have already been paid an extra attention to in the literature, our work demonstrates that one does not require gigantic data and compute resources for effective self-supervised pretraining. Such results promise a more democratized research arena where smaller groups are not alienated due lack of sufficient compute resources. More importantly, lowering the compute requirements naturally reduces any environmental effects which training these models can potentially have. ",
+        "bbox": [
+            165,
+            738,
+            825,
+            849
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Potential risks. Any machine learning method is susceptible to the potential underlying biases in the data. This is more important for self-supervised methods that deal with huge volumes, often not evaluated by diverse group of humans for any fairness concerns. The same is generally true in our case which requires us to make sure that titles that are included in training are diverse and inclusive. ",
+        "bbox": [
+            171,
+            856,
+            823,
+            911
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "387 References   \n388 [1] S. Abu-El-Haija, N. Kothari, J. Lee, P. Natsev, G. Toderici, B. Varadarajan, and S. Vijayanarasimhan.   \n389 Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016.   \n390 [2] J.-B. Alayrac, A. Recasens, R. Schneider, R. Arandjelovic, J. Ramapuram, J. De Fauw, L. Smaira, ´   \n391 S. Dieleman, and A. Zisserman. Self-Supervised MultiModal Versatile Networks. In NeurIPS, 2020.   \n392 [3] H. Alwassel, D. Mahajan, B. Korbar, L. Torresani, B. Ghanem, and D. Tran. Self-supervised learning by   \n393 cross-modal audio-video clustering. In Advances in Neural Information Processing Systems, volume 33,   \n394 pages 9758–9770, 2020.   \n395 [4] R. Arandjelovic and A. Zisserman. Look, listen and learn. In Proceedings of the IEEE International   \n396 Conference on Computer Vision, pages 609–617, 2017.   \n397 [5] P. Bachman, R. D. Hjelm, and W. Buchwalter. Learning representations by maximizing mutual information   \n398 across views. In Advances in Neural Information Processing Systems, pages 15535–15545, 2019.   \n399 [6] J. Carreira and A. Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In   \n400 proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299–6308,   \n401 2017.   \n402 [7] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual   \n403 representations. arXiv preprint arXiv:2002.05709, 2020.   \n404 [8] T. Chen, S. Kornblith, K. Swersky, M. Norouzi, and G. Hinton. Big self-supervised models are strong   \n405 semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020.   \n406 [9] C. Doersch, A. Gupta, and A. A. Efros. Unsupervised visual representation learning by context prediction.   \n407 In Proceedings of the IEEE international conference on computer vision, pages 1422–1430, 2015.   \n408 [10] B. Fernando, H. Bilen, E. Gavves, and S. Gould. Self-supervised video representation learning with   \n409 odd-one-out networks. In Proceedings of the IEEE conference on computer vision and pattern recognition,   \n410 pages 3636–3645, 2017.   \n411 [11] J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter.   \n412 Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE International Conference   \n413 on Acoustics, Speech and Signal Processing (ICASSP), pages 776–780. IEEE, 2017.   \n414 [12] D. Ghadiyaram, D. Tran, and D. Mahajan. Large-scale weakly-supervised pre-training for video action   \n415 recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages   \n416 12046–12055, 2019.   \n417 [13] S. Gidaris, P. Singh, and N. Komodakis. Unsupervised representation learning by predicting image   \n418 rotations. arXiv preprint arXiv:1803.07728, 2018.   \n419 [14] M. Gutmann and A. Hyvärinen. Noise-contrastive estimation: A new estimation principle for unnormalized   \n420 statistical models. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and   \n421 Statistics, pages 297–304, 2010.   \n422 [15] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick. Momentum contrast for unsupervised visual representation   \n423 learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages   \n424 9729–9738, 2020.   \n425 [16] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the   \n426 IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.   \n427 [17] O. J. Hénaff, A. Srinivas, J. De Fauw, A. Razavi, C. Doersch, S. Eslami, and A. v. d. Oord. Data-efficient   \n428 image recognition with contrastive predictive coding. arXiv preprint arXiv:1905.09272, 2019.   \n429 [18] R. D. Hjelm, A. Fedorov, S. Lavoie-Marchildon, K. Grewal, P. Bachman, A. Trischler, and Y. Bengio.   \n430 Learning deep representations by mutual information estimation and maximization. arXiv preprint   \n431 arXiv:1808.06670, 2018.   \n432 [19] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal   \n433 covariate shift. arXiv preprint arXiv:1502.03167, 2015.   \n434 [20] H. Kataoka, T. Wakamiya, K. Hara, and Y. Satoh. Would mega-scale datasets further enhance spatiotempo  \n435 ral 3d cnns? arXiv preprint arXiv:2004.04968, 2020. ",
+        "bbox": [
+            147,
+            69,
+            828,
+            921
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "36 [21] D. Kim, D. Cho, and I. S. Kweon. Self-supervised video representation learning with space-time cubic puzzles. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8545–8552, 2019.   \n39 [22] B. Korbar, D. Tran, and L. Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In Advances in Neural Information Processing Systems, pages 7763–7774, 2018. [23] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. Hmdb: a large video database for human motion recognition. In 2011 International Conference on Computer Vision, pages 2556–2563. IEEE, 2011.   \n43 [24] Z. Lai and W. Xie. Self-supervised learning for video correspondence flow. arXiv preprint arXiv:1905.00875, 2019.   \n5 [25] H.-Y. Lee, J.-B. Huang, M. Singh, and M.-H. Yang. Unsupervised representation learning by sorting sequences. In Proceedings of the IEEE International Conference on Computer Vision, pages 667–676, 2017. [26] D. Mahajan, R. Girshick, V. Ramanathan, K. He, M. Paluri, Y. Li, A. Bharambe, and L. van der Maaten. Exploring the limits of weakly supervised pretraining. In Proceedings of the European Conference on Computer Vision (ECCV), pages 181–196, 2018. [27] A. Miech, J.-B. Alayrac, L. Smaira, I. Laptev, J. Sivic, and A. Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879–9889, 2020. [28] A. Miech, D. Zhukov, J.-B. Alayrac, M. Tapaswi, I. Laptev, and J. Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE international conference on computer vision, pages 2630–2640, 2019. [29] I. Misra and L. v. d. Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707–6717, 2020. [30] I. Misra, C. L. Zitnick, and M. Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European Conference on Computer Vision, pages 527–544. Springer, 2016. [31] P. Morgado, N. Vasconcelos, and I. Misra. Audio-visual instance discrimination with cross-modal agreement. arXiv preprint arXiv:2004.12943, 2020.   \n63 [32] V. Nair and G. E. Hinton. Rectified linear units improve restricted boltzmann machines. In ICML, 2010.   \n4 [33] M. Noroozi and P. Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European Conference on Computer Vision, pages 69–84. Springer, 2016.   \n6 [34] M. Noroozi, H. Pirsiavash, and P. Favaro. Representation learning by learning to count. In Proceedings of the IEEE International Conference on Computer Vision, pages 5898–5906, 2017.   \n68 [35] A. v. d. Oord, Y. Li, and O. Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. [36] A. Owens and A. A. Efros. Audio-visual scene analysis with self-supervised multisensory features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631–648, 2018. [37] A. Owens, J. Wu, J. H. McDermott, W. T. Freeman, and A. Torralba. Ambient sound provides supervision for visual learning. In European conference on computer vision, pages 801–816. Springer, 2016. [38] D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, and A. A. Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536–2544, 2016. [39] M. Patrick, Y. M. Asano, R. Fong, J. F. Henriques, G. Zweig, and A. Vedaldi. Multi-modal self-supervision from generalized data transformations. arXiv preprint arXiv:2003.04298, 2020. [40] K. J. Piczak. Esc: Dataset for environmental sound classification. In Proceedings of the 23rd ACM international conference on Multimedia, pages 1015–1018, 2015. [41] S. Purushwalkam and A. Gupta. Demystifying contrastive self-supervised learning: Invariances, augmentations and dataset biases. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 3407–3418, 2020.   \n[42] K. Soomro, A. R. Zamir, and M. Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.   \n[43] C. Sun, A. Shrivastava, S. Singh, and A. Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843–852, 2017.   \n[44] Y. Tian, D. Krishnan, and P. Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849, 2019.   \n[45] D. Tran, H. Wang, L. Torresani, J. Ray, Y. LeCun, and M. Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450–6459, 2018.   \n[46] C. Vondrick, A. Shrivastava, A. Fathi, S. Guadarrama, and K. Murphy. Tracking emerges by colorizing videos. In Proceedings of the European conference on computer vision (ECCV), pages 391–408, 2018.   \n[47] F. Wang and H. Liu. Understanding the behaviour of contrastive loss. arXiv preprint arXiv:2012.09740, 2020.   \n[48] J. Wang, J. Jiao, L. Bao, S. He, Y. Liu, and W. Liu. Self-supervised spatio-temporal representation learning for videos by predicting motion and appearance statistics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4006–4015, 2019.   \n[49] D. Xu, J. Xiao, Z. Zhao, J. Shao, D. Xie, and Y. Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10334–10343, 2019.   \n[50] R. Zhang, P. Isola, and A. A. Efros. Colorful image colorization. In European conference on computer vision, pages 649–666. Springer, 2016. ",
+        "bbox": [
+            153,
+            98,
+            828,
+            916
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            161,
+            92,
+            826,
+            444
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Checklist ",
+        "text_level": 1,
+        "bbox": [
+            169,
+            468,
+            254,
+            483
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "1. For all authors... ",
+        "bbox": [
+            214,
+            494,
+            339,
+            508
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]   \n(b) Did you describe the limitations of your work? [Yes] Specifically, training data being proprietary creates concerns around reproducibility, which has precedence in the literature as mentioned in the paper. We address that partially by planning to publicly release pretrained models.   \n(c) Did you discuss any potential negative societal impacts of your work? [Yes]   \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ",
+        "bbox": [
+            236,
+            513,
+            825,
+            646
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "2. If you are including theoretical results... ",
+        "bbox": [
+            214,
+            651,
+            493,
+            665
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A] ",
+        "bbox": [
+            238,
+            669,
+            738,
+            700
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "3. If you ran experiments... ",
+        "bbox": [
+            214,
+            705,
+            393,
+            719
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No] The data is proprietary. However, we have provided implementation of the proposed method in supplemental material and aim to publicly release the pretrained models.   \n(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] They are all discussed in detail either in the main submission or in supplemental material.   \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] We only observed meaningful differences after running experiments multiple times, for ESC50[40] downstream experiments. Corresponding standard errors are reported in supplemental material.   \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] Please refer to Section 4.1. ",
+        "bbox": [
+            236,
+            724,
+            825,
+            911
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ",
+        "bbox": [
+            212,
+            90,
+            825,
+            107
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "(a) If your work uses existing assets, did you cite the creators? [N/A]   \n(b) Did you mention the license of the assets? [N/A]   \n(c) Did you include any new assets either in the supplemental material or as a URL? [N/A]   \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A]   \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A] ",
+        "bbox": [
+            236,
+            109,
+            825,
+            232
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "5. If you used crowdsourcing or conducted research with human subjects... ",
+        "bbox": [
+            212,
+            236,
+            705,
+            251
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A]   \n(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]   \n(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] ",
+        "bbox": [
+            238,
+            253,
+            826,
+            344
+        ],
+        "page_idx": 12
+    }
+]

parse/train/8hGabvaV2GQ/8hGabvaV2GQ_middle.json ADDED Viewed