ymyy307 commited on
Commit
b8671a6
·
verified ·
1 Parent(s): f0535fc

Upload additional_frame_merger_analysis.tex with huggingface_hub

Browse files
additional_frame_merger_analysis.tex ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ % Evidence-backed replacement using the deployed final selector checkpoint.
2
+ % This is an inference-time controlled ablation on a cached training subset,
3
+ % not the originally drafted two-model validation experiment.
4
+ \subsection{Additional Frame-Merger Analysis}
5
+ \label{supp:merger_analysis}
6
+
7
+ We additionally isolate how candidate redundancy affects the deployed
8
+ MiniMax-H3 selector. We evaluate the final selector checkpoint on all 404
9
+ locally cached training examples with at least six merged candidates, spanning
10
+ 378 videos. For the no-merger proxy, we reconstruct the 4\,fps candidate
11
+ stream by assigning each temporal sample to its nearest medoid within the same
12
+ shot; the sample inherits that medoid's feature and generator-relevance label.
13
+ Thus both settings use identical targets and selector weights, and differ only
14
+ in candidate redundancy. Succ.@5 is the fraction of examples for which the
15
+ hard Top-5 contains a candidate with maximum mean generator relevance.
16
+
17
+ Expanding the merged bank from 14.91 to 129.93 candidates leaves the ranking
18
+ loss nearly unchanged, but reduces NDCG@5 by 7.21 points and Succ.@5 by 49.01
19
+ points (Table~\ref{supp:tab:merger}). The result indicates that the main
20
+ failure mode is Top-5 slot competition among redundant views rather than a
21
+ large change in the average pairwise ranking objective.
22
+
23
+ \begin{table}[t]
24
+ \centering
25
+ \caption{Inference-time Frame-Merger ablation using the final MiniMax-H3
26
+ selector on a cached training subset (404 examples from 378 videos). Bank is
27
+ the average number of candidates.}
28
+ \label{supp:tab:merger}
29
+ \footnotesize
30
+ \setlength{\tabcolsep}{2.5pt}
31
+ \begin{tabular}{lcccc}
32
+ \toprule
33
+ Setting & Bank & $\mathcal{L}_{\mathrm{rank}}\downarrow$ & NDCG@5$\uparrow$ & Succ.@5$\uparrow$ \\
34
+ \midrule
35
+ 4\,fps-expanded proxy & 129.93 & 0.384 & 0.764 & 32.92\% \\
36
+ w/ Merger (Canon) & 14.91 & 0.378 & 0.836 & 81.93\% \\
37
+ \bottomrule
38
+ \end{tabular}
39
+ \end{table}
40
+
41
+ Table~\ref{supp:tab:merger_overhead} measures only latent-space similarity
42
+ grouping and medoid computation, after the input latents have already been
43
+ encoded and loaded. The production implementation performs this operation on
44
+ CPU, so it adds no GPU allocation. We report incremental peak CPU RSS above
45
+ the resident input bank. Short, Middle, and Long use cached examples nearest
46
+ the P10, P50, and P90 sizes of the reconstructed 4\,fps history distribution.
47
+
48
+ \begin{table}[t]
49
+ \centering
50
+ \caption{Frame-Merger overhead after latent encoding. Latency is the median
51
+ of 31 runs; peak memory is incremental CPU RSS above the resident input
52
+ latent bank. The merger has no learnable parameters and adds 0 MiB GPU
53
+ memory.}
54
+ \label{supp:tab:merger_overhead}
55
+ \footnotesize
56
+ \setlength{\tabcolsep}{3pt}
57
+ \begin{tabular}{lccccc}
58
+ \toprule
59
+ History & Frames in & Bank out & Latency & Peak mem. & Params \\
60
+ \midrule
61
+ Short & 48 & 4 & 7.0 ms & 11.9 MiB & 0 \\
62
+ Middle & 119 & 7 & 16.3 ms & 15.3 MiB & 0 \\
63
+ Long & 228 & 18 & 29.7 ms & 20.2 MiB & 0 \\
64
+ \bottomrule
65
+ \end{tabular}
66
+ \end{table}