Research map / Video / Action recognition
Decomposed Cross-Modal Distillation for RGB-Based Temporal Action Detection
Pilhyeon Lee, Taeoh Kim, Minho Shim, Dongyoon Wee, Hyeran Byun
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 |
|---|---|---|---|
| Use decomposed instead of conventional distillation (ablation) | TSM18 + Actionformer with temporal-gradient motion guidance · THUMOS'14 | average mAP (%): 44.5 -> 45.2 | Improved |
| Add conventional distillation (ablation) | RGB TSM18 + Actionformer · THUMOS'14 | average mAP (%): 43.6 -> 44.5 | Improved |
| Add decomposed distillation (ablation) | RGB TSM18 + Actionformer · THUMOS'14 | average mAP (%): 43.6 -> 45.2 | Improved |
| Add local attentive fusion to decomposed distillation (ablation) | TSM18 + Actionformer · THUMOS'14 | average mAP (%): 45.2 -> 46.6 (+1.4) | Improved |
| Replace local attentive fusion with summation (ablation) | Decomposed-distillation TSM18 + Actionformer · THUMOS'14 | average mAP (%): 46.6 -> 45.1 | Hurt |
| Replace local attentive fusion with self-attention (ablation) | Decomposed-distillation TSM18 + Actionformer · THUMOS'14 | average mAP (%): 46.6 -> 44.6 | Hurt |
| Replace local attentive fusion with cross-attention (ablation) | Decomposed-distillation TSM18 + Actionformer · THUMOS'14 | average mAP (%): 46.6 -> 44.2 | Hurt |
| Replace local attentive fusion with difference-attention (ablation) | Decomposed-distillation TSM18 + Actionformer · THUMOS'14 | average mAP (%): 46.6 -> 43.3 | Hurt |
| Distill motion knowledge from temporal gradients (ablation) | RGB-based detector, TSM18 + Actionformer · THUMOS'14 | average mAP (%): 43.6 -> 46.6 (+3.0) | Improved |
| Distill motion knowledge from optical flow (ablation) | RGB-based detector, TSM18 + Actionformer · THUMOS'14 | average mAP (%): 43.6 -> 48.2 (+4.6) | Improved |
| Use the proposed RGB-only distillation method instead of the RGB Actionformer baseline | Slowfast50 + Actionformer · THUMOS'14 | average mAP (%): 55.5 -> 56.8 | Improved |
| Use the proposed RGB-only distillation method instead of the RGB Actionformer baseline | Slowfast50 + Actionformer · ActivityNet1.3 val | average mAP: 34.94 -> 35.58 | Improved |
Closest papers by meaning
- Learning an Augmented RGB Representation With Cross-Modal Knowledge Distillation for Action Detection (ICCV 2021 · similarity 0.94)
- Graph Distillation for Action Detection with Privileged Modalities (ECCV 2018 · similarity 0.89)
- Modality Distillation with Multiple Stream Networks for Action Recognition (ECCV 2018 · similarity 0.88)
- End-to-end Multi-modal Video Temporal Grounding (NeurIPS 2021 · similarity 0.88)
- CNN-Based RGB-D Salient Object Detection: Learn, Select, and Fuse (IJCV 2021 · similarity 0.87)
- Local-Global Multi-Modal Distillation for Weakly-Supervised Temporal Video Grounding (AAAI 2024 · similarity 0.86)
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.
