Nuntea/vjepa2-temporal-order-blindspots

V-JEPA2 Temporal Order Blind Spots

This dataset documents blind spots for the pretrained video world model facebook/vjepa2-vith-fpc64-256.

The model produces nearly identical embeddings for videos whose temporal order has been severely corrupted, indicating weak sensitivity to temporal directionality and causal motion structure.

Model Tested

Model name: facebook/vjepa2-vith-fpc64-256
Type: Self-supervised video world model (JEPA-style joint embedding predictive architecture)
Modality: Video
Model card: https://huggingface.co/facebook/vjepa2-vith-fpc64-256

This is a base pretrained model, not fine-tuned for classification, captioning, or action recognition.

Evaluation Setup

Dataset

• Huggingface-ID: nateraw/kinetics-mini
• Source videos: Kinetics-Mini (validation split)
• Each video was transformed to simulate temporal corruption.

Temporal Transformations Tested

Each original video was transformed in ways that should change its semantic meaning:

Transformation	Description
full_reverse	Entire video reversed in time
block_reverse	Each block of 8 frames reversed
pingpong	Forward playback followed by backward playback

Metric

• Cosine similarity between original and transformed video embeddings

Observed Blind Spot

Despite drastic temporal corruption, the model outputs very high cosine similarity:

• Typical similarity range: 0.97 – 0.995 (one case of 0.94)

This indicates that the model:

• Treats time-reversed motion as equivalent
• Fails to encode the notion of time
• Is largely invariant to local temporal order

For many actions (e.g., throwing, opening, jumping), reversing time should change the meaning — but the embeddings remain nearly unchanged.

Blind Spots Dataset Structure

vjepa2-temporal-order-blindspots/

• metadata.csv: Contains video IDs, applied transformations, and similarity scores.
• videos/: Stores all video files. For each original video, multiple transformed versions are included:
  • Original video (*_orig.mp4)
  • Fully reversed video (*_full_reverse.mp4)
  • Block-wise reversed video, with each block of 8 frames reversed (*_block_reverse.mp4)
  • Ping-pong video, which plays forward then backward (*_pingpong.mp4)

How to Load the Dataset

from datasets import load_dataset

dataset = load_dataset("Nuntea/vjepa2-temporal-order-blindspots", split="train")

row = dataset[0]
print(row)
print(row["class_id"], row["transformation"], row["cosine_similarity"])

video = row["video"] 

Each row contains:

• video
• Class label
• Transformation type
• Precomputed Cosine similarity metric with respect to the original video.

Metadata Description

Column	Description
file_name	Transformed video filename
class_id	Kinetics class ID
transformation	Name of transformation
transformation_description	Human-readable description
cosine_similarity	Similarity to original embedding

Expected vs Actual Output

Expected behavior:

• Temporal transformations should deivate and reduce embedding similarity, especially for full_reverse.

Actual behavior:

• The model produces embeddings nearly invariant to temporal order.

Each row in this dataset represents a failure case where the model’s output does not match the expected semantic distinction.

Why This Happens (My Hypothesis)

V-JEPA-style training emphasizes:

• Predictive consistency
• Spatial semantics
• Appearance invariance

However, it does not explicitly penalize temporal inversion, leading to:

• Weak causal modeling
• Motion treated as unordered frame sets
• Poor sensitivity to temporal directionality

How This Could Be Fixed with Fine-Tuning

The blind spot in V-JEPA2 arises because the model treats time-reversed or temporally scrambled videos almost identically. To address this, the model should be fine-tuned with a temporal discrimination objective that explicitly teaches it to recognize the direction and order of motion.

Recommended Dataset

The dataset should contain paired videos emphasizing temporal structure:

• Forward vs backward: original video and its reversed version
• Scrambled vs coherent: clips with frame order shuffled vs normal
• Causal vs anti-causal: actions that have clear cause-effect order (e.g., "throwing a ball" vs "ball flying into hand")

Example:

Original	Transformed	Label
Person throws ball	Ball flies back into hand	Backward
Pouring water	Frames shuffled	Scrambled
Door opens	Door closes in reverse	Anti-causal

Possible Data Sources

Existing video datasets:

• Kinetics / Something-Something (apply synthetic temporal transformations: reverse, ping-pong, block-reverse, shuffle)

Physics-based or procedural motion datasets:

• Synthetic videos of moving objects or simulations

Procedurally generated motion sequences for causal tasks

• Augmentation: Temporal jittering, frame dropping, slow-motion / fast-forward to encourage temporal sensitivity

Training Objective

A potential fine-tuning objective should explicitly reward correct temporal ordering:

Contrastive Learning:

• Anchor: original clip
• Positive: same clip in forward order
• Negative: reversed or scrambled clip
• Loss: InfoNCE or cosine similarity loss

Temporal Classification Head:

• Predict whether the video is forward, backward, or scrambled
• Direction-Aware Predictive Modeling:
• Predict next frames conditioned on current sequence (like autoregressive temporal modeling)
• Penalize physically impossible sequences (anti-causal)

Estimated Dataset Size

• Small-scale proof-of-concept: ~5k–10k video pairs
• Full fine-tuning: 20k–50k+ video pairs
• Using strong negative examples (reversed / scrambled / anti-causal) allows smaller datasets to be effective

Intended Use

This dataset is intended for:

• Model diagnostics
• Video world model evaluation

It is not a classification dataset.