Research map / Video / Action recognition
Spatio-Temporal Crop Aggregation for Video Representation Learning
Sepehr Sameni, Simon Jenni, Paolo Favaro
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 SCALE aggregation of clip features on top of frozen backbone (linear probe) | SVT backbone · SSv2 | top-1 accuracy: 20.30 -> 25.26 | Improved |
| add SCALE aggregation (linear probe) on Kinetics-400 | SVT, rhoBYOL, VMAE · Kinetics-400 | top-1 accuracy: 71.71 -> 71.78 (SVT) | No clear effect |
| initialize transformer head with SCALE pretraining instead of random init | SCALE_ft · Kinetics-400 (SVT) | top-1 accuracy: 72.18 -> 72.38 | Improved |
| SCALE k-NN probing on MAE-based VMAE-B | VMAE-B · UCF-101 | k-NN accuracy: 35.05 -> 65.63 | Improved |
| SCALE k-NN probing on rhoBYOL | rhoBYOL · UCF-101 | k-NN accuracy: 85.19 -> 83.47 | Hurt |
| SCALE_ft in low-shot 10% Kinetics-400 | SCALE_ft · Kinetics-400 10% | top-1 accuracy: 66.43 -> 67.01 (SVT linear vs SCALE_ft) | Improved |
| add SCALE to long-form video (LVU Relation, SVT) | SVT · LVU Relation | accuracy: 64.70 -> 76.47 | Improved |
| use both SET and MCM losses instead of one (ablation) | SCALE_ft · UCF-101 (SVT) | accuracy: 91.80 -> 93.20 | Improved |
| masking ratio 0.25 versus 0.45 (ablation) | SCALE_ft · HMDB-51 (SVT) | accuracy: 63.39 -> 65.49 | Improved |
| use 2 transformer layers instead of 1 (ablation) | SCALE_ft · HMDB-51 (rhoBYOL) | accuracy: 64.39 -> 65.62 | Improved |
| increase views per set from 4 to 8 (ablation) | SCALE_ft · UCF-101 (SVT, batch 512) | accuracy: 92.67 -> 93.20 | Improved |
| increase batch size from 512 to 1024 (ablation) | SCALE_ft · UCF-101 (SVT, 6x2 views) | accuracy: 93.18 -> 92.96 | Hurt |
Closest papers by meaning
- Space-Time Crop & Attend: Improving Cross-Modal Video Representation Learning (ICCV 2021 · similarity 0.90)
- Self-Supervised Spatio-Temporal Representation Learning for Videos by Predicting Motion and Appearance Statistics (CVPR 2019 · similarity 0.86)
- ActionVLAD: Learning Spatio-Temporal Aggregation for Action Classification (CVPR 2017 · similarity 0.86)
- A Large-Scale Study on Unsupervised Spatiotemporal Representation Learning (CVPR 2021 · similarity 0.86)
- Self-Supervised Spatiotemporal Learning via Video Clip Order Prediction (CVPR 2019 · similarity 0.86)
- Compositional Video Understanding with Spatiotemporal Structure-based Transformers (CVPR 2024 · similarity 0.84)
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.
