Research map / Video / Action recognition
RESOUND: Towards Action Recognition without Representation Bias
Yingwei Li, Yi Li, Nuno Vasconcelos
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 |
|---|---|---|---|
| choose 3 classes by explicit RESOUND optimization instead of random sampling | 3-class dataset sampling · KTH (K=3 subset) | object bias (lower is better): 0.99 ± 0.09 -> 0.39 | Improved |
| choose 3 classes by explicit RESOUND optimization instead of random sampling | 3-class dataset sampling · Hollywood2 (K=3 subset) | object bias (lower is better): 0.86 ± 0.07 -> 0.44 | Improved |
| choose 3 classes by explicit RESOUND optimization instead of random sampling | 3-class dataset sampling · HMDB51 (K=3 subset) | object bias (lower is better): 0.82 ± 0.54 -> 0.00 | Improved |
| choose 3 classes by explicit RESOUND optimization instead of random sampling | 3-class dataset sampling · UCF101 (K=3 subset) | object bias (lower is better): 1.08 ± 0.02 -> 0.55 | Improved |
| choose 3 classes by explicit RESOUND optimization instead of random sampling | 3-class dataset sampling · ActivityNet (K=3 subset) | object bias (lower is better): 0.89 ± 0.10 -> 0.41 | Improved |
| choose 3 classes by explicit RESOUND optimization instead of random sampling | 3-class dataset sampling · Kinetics (K=3 subset) | object bias (lower is better): 1.00 ± 0.11 -> 0.41 | Improved |
| choose 3 classes by explicit RESOUND optimization instead of random sampling | 3-class dataset sampling · Charades (K=3 subset) | object bias (lower is better): 0.62 ± 0.20 -> 0.00 | Improved |
| increase temporal extent of C3D from 8 to 64 frames (ablation) | C3D · Diving48 | accuracy (%): 11.51 -> 27.60 | Improved |
| add optical flow stream to TSN RGB | TSN · Diving48 | accuracy (%): 16.77 -> 20.28 | Improved |
| use Diving48 instead of Kinetics as dataset | ResNet50 static features · representation bias (object) | B(D, phi_object): 4.51 -> 1.48 | Improved |
| increase dataset size (Kinetics vs smaller datasets) | action recognition datasets · object bias | B(D, phi_object): 3.16 -> 4.51 | Hurt |
Closest papers by meaning
- Enhancing Unsupervised Video Representation Learning by Decoupling the Scene and the Motion (AAAI 2021 · similarity 0.84)
- Discriminative Video Representation Learning Using Support Vector Classifiers (TPAMI 2019 · similarity 0.84)
- Exploring the Camera Bias of Person Re-identification (ICLR 2025 · similarity 0.83)
- FineParser: A Fine-grained Spatio-temporal Action Parser for Human-centric Action Quality Assessment (CVPR 2024 · similarity 0.82)
- Bias for Action: Video Implicit Neural Representations with Bias Modulation (CVPR 2025 · similarity 0.82)
- ALBAR: Adversarial Learning approach to mitigate Biases in Action Recognition (ICLR 2025 · similarity 0.82)
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.
