Research map / Video / Video segmentation and tracking
ViC-MAE: Self-Supervised Representation Learning from Images and Video with Contrastive Masked Autoencoders
Jefferson Hernandez, Ruben Villegas, Vicente Ordonez
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| sample frames with distant sampling instead of the same frame (ablation) | ViC-MAE · ImageNet-1K linear eval (ViT-B/16) | top-1 accuracy: 63.25 -> 67.66 | Improved |
| use mean pooling instead of GeM or max pooling for global representation (ablation) | ViC-MAE · ImageNet-1K linear eval (ViT-B/16) | top-1 accuracy: 66.92 -> 67.66 | Improved |
| combine color and spatial augmentation instead of color only (ablation) | ViC-MAE · ImageNet-1K linear eval (ViT-B/16) | top-1 accuracy: 65.40 -> 67.66 | Improved |
| pool local features for contrastive loss instead of using CLS token with SimSiam (ablation) | MAE + contrastive · ImageNet-1K linear eval (ViT-B/16) | top-1 accuracy: 58.58 -> 67.66 | Improved |
| pool local features for contrastive loss instead of using CLS token with VicReg (ablation) | MAE + contrastive · ImageNet-1K linear eval (ViT-B/16) | top-1 accuracy: 63.86 -> 67.66 | Improved |
| add strong color augmentation on top of strong spatial augmentation for video frames (ablation) | ViC-MAE · ImageNet-1K linear eval (ViT-B/16) | top-1 accuracy: 66.03 -> 67.66 | No clear effect |
| add contrastive objective to masked autoencoder (ViC-MAE vs OmniMAE on same data) | ViT-L/16 K400+IN1K · Kinetics-400 finetune | top-1 accuracy: 84.0 -> 86.8 | Improved |
| add contrastive objective to MAE in video-to-image transfer (ImageNet) | ViT-L/16 MiT only · ImageNet-1K finetune (video-only pretraining) | top-1 accuracy: 84.7 -> 85.3 | Improved |
| ViC-MAE versus supervised TubeViT on SSv2 | ViC-MAE · Something-Something v2 finetune (out-of-domain) | top-1 accuracy: 76.1 -> 75.9 | No clear effect |
Closest papers by meaning
- Contrastive Tuning: A Little Help to Make Masked Autoencoders Forget (AAAI 2024 · similarity 0.90)
- A Theoretical Analysis of Self-Supervised Learning for Vision Transformers (ICLR 2025 · similarity 0.89)
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.89)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.88)
- Self-Supervised Representation Learning from Arbitrary Scenarios (CVPR 2024 · similarity 0.88)
- Hybrid Distillation: Connecting Masked Autoencoders with Contrastive Learners (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
