Research map / Video / Video segmentation and tracking
How Severe Is Benchmark-Sensitivity in Video Self-Supervised Learning?
Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, Cees G. M. Snoek
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 |
|---|---|---|---|
| add self-supervised video pre-training before finetuning, on UCF-101 | R(2+1)D-18 · UCF-101 (full finetuning) | top-1 accuracy: 77.3 -> 91.3 (GDT) | Improved |
| use image-based MoCo pre-training instead of no pre-training on SS-v2 and EK-100 | R(2+1)D-18 · SS-v2 / EPIC-Kitchens-100 (finetuning) | top-1 accuracy: 57.1 -> 57.1 (SS-v2), 25.7 -> 26.4 (EK-100) | No clear effect |
| use supervised Kinetics pre-training instead of the best self-supervised method | R(2+1)D-18 · UCF-101 (full finetuning) | top-1 accuracy: 91.3 -> 93.9 | Improved |
| use supervised pre-training instead of the best self-supervised method under a large domain shift | R(2+1)D-18 · EPIC-Kitchens-100 (full finetuning) | top-1 accuracy: 42.8 -> 47.7 | Improved |
| use supervised pre-training instead of best self-supervised for linear evaluation | R(2+1)D-18 · Kinetics-400 (linear evaluation) | top-1 accuracy: 46.0 -> 65.9 | Improved |
| use contrastive AVID-CMA instead of supervised pre-training in linear evaluation on NTU-60 | R(2+1)D-18 · NTU-60 (linear evaluation) | top-1 accuracy: 45.5 -> 53.9 | Improved |
| reduce finetuning samples to 1000 on UCF-101 | self-supervised video methods · UCF-101 (low data) | top-1 accuracy | Hurt |
| use CtP instead of the second-best self-supervised method on FX-S1 | R(2+1)D-18 · FineGym FX-S1 | mean per-class accuracy: 68.0 -> 79.1 (12%) | Improved |
| use MoCo instead of VideoMoCo (frame dropout invariance) on fine-grained Gym-99 subsets | R(2+1)D-18 · FineGym FX-S1 | mean per-class accuracy: 57.3 -> 65.0 | Improved |
| switch from balanced Gym-99 to long-tailed Gym-288 for supervised pre-training | R(2+1)D-18 · FineGym | top-1 accuracy: 88.6 -> 58.4 | Hurt |
| add self-supervised pre-training for spatio-temporal action detection on UCF-101 | R(2+1)D-18 · UCF-101 action detection | mAP: 0.327 -> 0.476 (TCLR) | Improved |
| use supervised pre-training instead of self-supervised on Charades multi-label classification | R(2+1)D-18 · Charades | mAP: 12.2 -> 23.5 | Improved |
Closest papers by meaning
- Scaling and Benchmarking Self-Supervised Visual Representation Learning (ICCV 2019 · similarity 0.89)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.88)
- How Well Do Self-Supervised Models Transfer? (CVPR 2021 · similarity 0.87)
- Uncovering the Hidden Dynamics of Video Self-supervised Learning under Distribution Shifts (NeurIPS 2023 · similarity 0.87)
- How Useful Is Self-Supervised Pretraining for Visual Tasks? (CVPR 2020 · similarity 0.87)
- Is ImageNet worth 1 video? Learning strong image encoders from 1 long unlabelled video (ICLR 2024 · similarity 0.85)
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.
