Research map / Video / Action recognition
Learning End-to-end Video Classification with Rank-Pooling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| end-to-end training with rank-pooling layer (vs. non-end-to-end rank-pool + SVM) (ablation) | CNN temporal pooling on UCF-sports · UCF-sports | accuracy (%): 66 -> 87 (about 21 points) | Helped |
| end-to-end training with rank-pooling layer (vs. non-end-to-end rank-pool) (ablation) | CNN temporal pooling on Hollywood2 · Hollywood2 | mAP (%): 31.0 -> 40.6 (+9.6 mAP) | Helped |
| end-to-end training for avg and max pooling (ablation) | CNN temporal pooling on UCF-sports · UCF-sports | accuracy (%): avg 67->70, max 66->71 | Helped |
| diagonal approximation of rank-pool gradient instead of full gradient (ablation) | end-to-end rank-pool CNN · UCF-sports | accuracy (%): 87 vs 87 | No clear effect |
| end-to-end rank-pool CNN with full gradient vs. diagonal approximation, training speed (ablation) | end-to-end rank-pool CNN · training throughput (Titan-X) | frames per second: full gradient 5 fps vs diagonal 50 fps | Hurt |
| frame-level fine-tuning on top of rank-pooling (fine-tuned features then rank-pooled) (ablation) | Hollywood2 CNN features · Hollywood2 | mAP (%): 34.1 -> 36.3 mAP | Helped |
| end-to-end rank-pool training vs frame-level fine-tuning with rank-pooling (ablation) | Hollywood2 · Hollywood2 | mAP (%): 36.3 -> 40.6 | Helped |
| rank-pooling vs max/avg pooling without end-to-end training | CNN features with SVM · UCF-sports | accuracy (%): 66-67 across pooling methods | No clear effect |
| end-to-end rank-pool CNN vs avg/max pooling end-to-end on Hollywood2 | Hollywood2 · Hollywood2 | mAP (%): 40.6 vs avg+CNN 37.9 and max+CNN 34.7; avg+SVM 35.7 | Mixed |
| adding end-to-end-trained CNN features to HOG+HOF+MBH (ablation) | Hollywood2 hand-crafted features · Hollywood2 | mAP (%): 71.4 -> 73.4 mAP | Helped |
Closest papers by meaning
- Discriminative Hierarchical Rank Pooling for Activity Recognition (CVPR 2016 · similarity 0.92)
- Video Representation Learning Using Discriminative Pooling (CVPR 2018 · similarity 0.91)
- Generalized Rank Pooling for Activity Recognition (CVPR 2017 · similarity 0.90)
- Dynamic Image Networks for Action Recognition (CVPR 2016 · similarity 0.89)
- AdaScan: Adaptive Scan Pooling in Deep Convolutional Neural Networks for Human Action Recognition in Videos (CVPR 2017 · similarity 0.89)
- Beyond Short Clips: End-to-End Video-Level Learning With Collaborative Memories (CVPR 2021 · similarity 0.89)
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.