Research map / Continual learning / Class-incremental
Class-Incremental Learning for Action Recognition in Videos
Jaeyoo Park, Minsoo Kang, Bohyung Han
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 time-channel importance map to the distillation loss (ablation) | TCD · UCF101 (5x10 stages) | average incremental accuracy (CNN): 71.21 -> 72.61 | Improved |
| add orthogonality loss to distillation without importance map (ablation) | TCD · UCF101 (5x10 stages) | average incremental accuracy (CNN): 71.21 -> 72.31 | Improved |
| add orthogonality loss on top of importance-map distillation (ablation) | TCD · UCF101 (5x10 stages) | average incremental accuracy (CNN): 72.61 -> 73.43 | Improved |
| add orthogonality loss on top of importance-map distillation (NME) (ablation) | TCD · UCF101 (5x10 stages) | average incremental accuracy (NME): 74.81 -> 75.35 | Improved |
| add orthogonality loss to distillation (ablation) | TCD · UCF101 (10x5 stages, vs FT/EWC table) | average incremental accuracy (CNN): 73.09 -> 74.89 | Improved |
| replace EWC with TCD distillation | fine-tuning with exemplars · UCF101 (10x5 stages) | average incremental accuracy (CNN): 69.70 -> 74.89 | Improved |
| increase memory per class from 1 to 10 (ablation) | TCD · UCF101 (10 steps, memory budget) | average incremental accuracy (CNN): 64.52 -> 76.66 | Improved |
| store all frames of each exemplar video rather than uniform sampling (ablation) | TCD · UCF101 (10 steps, sampling) | average incremental accuracy (CNN): 73.43 -> 73.89 | No clear effect |
| random frame sampling instead of uniform (ablation) | TCD · UCF101 (10 steps, sampling) | average incremental accuracy (CNN): 73.43 -> 73.17 | No clear effect |
| increase input frames from 8 to 16 (ablation) | TCD · UCF101 (10 stages, frames) | average incremental accuracy (CNN): 73.43 -> 75.17 | Improved |
| decrease input frames from 8 to 1 (ablation) | TCD · UCF101 (10 stages, frames) | average incremental accuracy (CNN): 73.43 -> 69.14 | Hurt |
| combine TCD distillation with BiC | BiC · UCF101 (2x25 stages) | average incremental accuracy (CNN): 68.85 -> 72.00 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- iCaRL — iCaRL: Incremental Classifier and Representation Learning (CVPR 2017) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- vCLIMB: A Novel Video Class Incremental Learning Benchmark (CVPR 2022 · similarity 0.91)
- Class-Incremental Learning: Survey and Performance Evaluation on Image Classification (TPAMI 2022 · similarity 0.89)
- Learning a Condensed Frame for Memory-Efficient Video Class-Incremental Learning (NeurIPS 2022 · similarity 0.89)
- Continual Learning Using a Kernel-Based Method Over Foundation Models (AAAI 2025 · similarity 0.87)
- Hypercorrelation Evolution for Video Class-Incremental Learning (AAAI 2024 · similarity 0.87)
- Revisiting Generative Replay for Class Incremental Object Detection (CVPR 2025 · 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.
