Research map / Video / Action recognition
Uncertainty-Aware Weakly Supervised Action Detection from Untrimmed Videos
Anurag Arnab, Chen Sun, Arsha Nagrani, Cordelia Schmid
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 multiple instance learning (max pooling) over tubelets instead of using video label per tubelet (ablation) | weakly supervised action detector · UCF101-24 val | Video AP@0.5: 29.7 -> 33.5 | Improved |
| add uncertainty-based loss to MIL max pooling (ablation) | MIL max pooling · UCF101-24 val | Video AP@0.5: 33.5 -> 35.0 | Improved |
| add uncertainty-based loss to MIL max pooling (ablation) | MIL max pooling · UCF101-24 val | Video AP@0.2: 60.7 -> 61.7 | Improved |
| use max pooling instead of mean pooling (ablation) | MIL · UCF101-24 val | Video AP@0.5: 33.0 -> 33.5 | Improved |
| use max pooling instead of LSE pooling (ablation) | MIL · UCF101-24 val | Video AP@0.5: 33.1 -> 33.5 | Improved |
| aggregation function choice (LSE, mean, max) (ablation) | MIL · UCF101-24 val | Video AP@0.2: 60.1 / 60.3 / 60.7 | No clear effect |
| add uncertainty loss with 3 bags x 5 tubelets (ablation) | MIL max pooling · UCF101-24 val | Video AP@0.5: 33.6 -> 34.1 | Improved |
| add uncertainty loss with 2 bags x 8 tubelets (ablation) | MIL max pooling · UCF101-24 val | Video AP@0.5: 33.3 -> 34.2 | Improved |
| add uncertainty loss with 1 bag x 16 tubelets (ablation) | MIL max pooling · UCF101-24 val | Video AP@0.5: 25.8 -> 26.2 | Improved |
| use fewer, larger bags (1 bag of 16 tubelets instead of 4 bags of 4) (ablation) | MIL max pooling + uncertainty · UCF101-24 val | Video AP@0.5: 35.0 -> 26.2 | Hurt |
| shorten the sub-clip duration of weak labels from 900 s to 1 s | MIL max pooling + uncertainty · AVA v2.2 | Frame AP@0.5: 4.2 -> 22.4 | Improved |
| full box supervision instead of weak 5 s sub-clip labels | SlowFast R50 · AVA v2.2 | Frame AP@0.5: 18.0 -> 24.9 | Improved |
Closest papers by meaning
- Weakly-supervised Temporal Action Localization by Uncertainty Modeling (AAAI 2021 · similarity 0.91)
- Large-Scale Weakly-Supervised Pre-Training for Video Action Recognition (CVPR 2019 · similarity 0.91)
- A flexible model for training action localization with varying levels of supervision (NeurIPS 2018 · similarity 0.91)
- Multi-Instance Multi-Label Action Recognition and Localization Based on Spatio-Temporal Pre-Trimming for Untrimmed Videos (AAAI 2020 · similarity 0.90)
- Activity Driven Weakly Supervised Object Detection (CVPR 2019 · similarity 0.90)
- Stable Mean Teacher for Semi-supervised Video Action Detection (AAAI 2025 · similarity 0.88)
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.
