Research map / Video / Action recognition
Hierarchical Relational Networks for Group Activity Recognition and Retrieval
Mostafa S. Ibrahim, Greg Mori
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 a relational layer (all-pairs) in place of the shared dense layer (ablation) | B1-NoRelations · Volleyball group activity (single frame) | accuracy: 85.1 -> 86.5 | Improved |
| fine-tune VGG19 on person boxes (ablation) | RCRG-1R-1C · Volleyball group activity (single frame) | accuracy: 75.4 -> 86.5 | Improved |
| split first layer into 2 team cliques instead of all-people clique (ablation) | RCRG-2R · Volleyball group activity (single frame) | accuracy: 86.1 -> 87.2 | Improved |
| stack a third relational layer (4,2,1 cliques) over 2 layers (ablation) | RCRG-2R-21C · Volleyball group activity (single frame) | accuracy: 87.2 -> 86.4 | Hurt |
| use concatenation pooling instead of max-pooling (ablation) | RCRG-2R-11C · Volleyball group activity (single frame) | accuracy: 86.1 -> 88.3 | Improved |
| use concatenation pooling instead of max-pooling (ablation) | RCRG-2R-21C · Volleyball group activity (single frame) | accuracy: 87.2 -> 86.7 | Hurt |
| use concatenation pooling instead of max-pooling (ablation) | RCRG-3R-421C · Volleyball group activity (single frame) | accuracy: 86.4 -> 87.3 | Improved |
| use relational autoencoder instead of per-person compact autoencoder | B1-Compact128 · Volleyball scene retrieval | Hit@1: 49.4 -> 57.4 | Improved |
| use relational autoencoder (128-d) instead of raw VGG19 features (4096-d) | B2-VGG19 · Volleyball scene retrieval | Hit@1: 55.0 -> 57.4 | Improved |
| use 4 layers with 2 team cliques instead of 2 layers with 1 clique (ablation) | RAER · Volleyball scene retrieval (model variants) | Hit@1: 56.8 -> 57.4 | Improved |
| use hierarchical cliques (4,2,2,4) instead of uniform 2 cliques (ablation) | RAER-4L · Volleyball scene retrieval (model variants) | Hit@1: 57.4 -> 55.8 | Hurt |
| use relational autoencoder instead of VGG19 when 50% of people are missing | B2-VGG19-D · Volleyball scene retrieval (denoising, 50% people dropped) | Hit@1: 34.0 -> 43.0 | Improved |
Closest papers by meaning
- Empowering Relational Network by Self-Attention Augmented Conditional Random Fields for Group Activity Recognition (ECCV 2020 · similarity 0.90)
- Convolutional Relational Machine for Group Activity Recognition (CVPR 2019 · similarity 0.89)
- Learning Actor Relation Graphs for Group Activity Recognition (CVPR 2019 · similarity 0.89)
- Structure Inference Machines: Recurrent Neural Networks for Analyzing Relations in Group Activity Recognition (CVPR 2016 · similarity 0.87)
- Detector-Free Weakly Supervised Group Activity Recognition (CVPR 2022 · similarity 0.87)
- Group Activity Prediction with Sequential Relational Anticipation Model (ECCV 2020 · similarity 0.87)
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.
