Research map / Video / Action recognition
Weakly-Supervised Action Localization, and Action Recognition Using Global–Local Attention of 3D CNN
Novanto Yudistira, Muthu Subash Kavitha, Takio Kurita
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 |
|---|---|---|---|
| aggregate 3DCAM over all layers instead of GradCAM on last layer | 3D ResNeXt-101 · JHMDB pointing game | accuracy: 0.988 -> 0.991 | No clear effect |
| use Sum 3DCAM instead of GradCAM | 3D ResNeXt-101 · JHMDB localization IoU 0.5 and 0.6 | localization accuracy: 58.75 % and 284.61 % improvement at IoU 0.5 and 0.6 | Improved |
| use Sum 3DCAM instead of GradCAM for joint classification and localization | 3D ResNeXt-101 · JHMDB Acc2 | Acc2: 46.15 % and 90.0 % improvement at IoU 0.5 and 0.6 | Improved |
| sum aggregation instead of average pooling of layer 3DCAMs (ablation) | 3DCAM · JHMDB localization | localization accuracy: 0.089 -> 0.127 at IoU 0.5 | Improved |
| use only single layers (first or fourth) rather than all layers (ablation) | 3DCAM on JHMDB · JHMDB pointing game | accuracy: 0.991 -> 0.821 | Hurt |
| aggregate conv and layer 1 instead of conv layer alone (ablation) | 3DCAM · UCF24 | MaxBoxAcc average: 0.159 -> 0.395 | Improved |
| aggregate layers conv and 1 instead of a single layer 1 (ablation) | 3DCAM · UCF24 | MaxBoxAcc average: 0.366 -> 0.395 | Improved |
| add layers 2 and 3 to the conv+layer1 aggregation (ablation) | Sum 3DCAM · UCF24 | MaxBoxAcc average: 0.395 -> 0.37 | Hurt |
| use 3DCAM instead of GradCAM with same layer composition | Sum aggregation (conv and layer 1) · UCF24 | MaxBoxAcc average: 0.361 -> 0.395 (2.5% difference stated) | Improved |
| aggregate 3DCAM layers 2,3,4 instead of GradCAM layer 4 only | ResNet50 · ILSVRC localization | MaxBoxAcc average: 0.57 -> 0.598 | Improved |
| 3DCAM sum instead of Sum GradCAM at IoU 0.5 | ResNet50 · ILSVRC localization | MaxBoxAcc IoU 0.5: 0.622 -> 0.622 | No clear effect |
| add layer-wise attention gates to 3D ResNeXt-101 | 3D ResNeXt-101 · HMDB51 | video-wise accuracy: 0.891 -> 0.919 | Improved |
Closest papers by meaning
- Temporal Action Localization in Untrimmed Videos via Multi-Stage CNNs (CVPR 2016 · similarity 0.91)
- Deep Analysis of CNN-Based Spatio-Temporal Representations for Action Recognition (CVPR 2021 · similarity 0.90)
- Temporal Action Localization in the Deep Learning Era: A Survey (TPAMI 2023 · similarity 0.90)
- T-C3D: Temporal Convolutional 3D Network for Real-Time Action Recognition (AAAI 2018 · similarity 0.90)
- 3C-Net: Category Count and Center Loss for Weakly-Supervised Action Localization (ICCV 2019 · similarity 0.89)
- Graph Convolutional Module for Temporal Action Localization in Videos (TPAMI 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.
