Research map / Probabilistic and scientific modeling / Causal discovery
Bi-Causal: Group Activity Recognition via Bidirectional Causality
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Bi-Causal, a group activity recognition network that models human relations and human-object interactions with a bidirectional causality communication channel and KL-divergence consistency, reporting SOTA on Volleyball and Collective Activity.
Problem
Group activity recognition has focused on human relations while neglecting human-object interactions and the bidirectional dependence between relations and interactions.
Approach
Uses Granger causality tests to show bidirectional causality between human relations (HRs) and human-object interactions (HOIs), then builds an Interaction Module (graph + temporal self-attention over object and person nodes) and a Causality Communication Channel with exchange tokens and symmetric KL-divergence between relation and interaction class scores.
Key results
- Bi-Causal reaches 96.1% group activity accuracy on VOLLEYBALL, versus 94.1% for GroupFormer-style COMPOSER at 94.6%, a state-of-the-art result among keypoint-only methods on that dataset.
- On RE-ANNOTATED VOLLEYBALL, Bi-Causal achieves 96.8% accuracy, above COMPOSER (96.2%) and GroupFormer (94.4%).
- Ablation shows bidirectional causality raises accuracy from 93.1% (RM only) to 96.1% on VOLLEYBALL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add bidirectional causality between HRs and HOIs (ablation) | Bi-Causal · VOLLEYBALL | accuracy: 93.1 -> 96.1 | Helped |
| Add HR-to-HOI causality (ablation) | Interaction Module · VOLLEYBALL | accuracy: 94.7 -> 95.4 | Helped |
| Add HOI-to-HR causality (ablation) | Relation Module · VOLLEYBALL | accuracy: 93.1 -> 94.3 | Helped |
| Use bidirectional causality instead of combining HRs and HOIs without causality (ablation) | Bi-Causal · VOLLEYBALL | accuracy: 94.7 -> 96.1 | Helped |
| Use a GCN Interaction Module instead of a feed-forward network (ablation) | Interaction Module · VOLLEYBALL | accuracy: 91.9 -> 95.3 | Helped |
| Add ball information (ablation) | Interaction Module · VOLLEYBALL | accuracy: 92.7 -> 95.3 | Helped |
| Add self-attention over graph edges (ablation) | Interaction Module · VOLLEYBALL | accuracy: 93.6 -> 95.3 | Helped |
| Combine spatial-temporal and temporal-spatial relation paths (ablation) | Relation Module · VOLLEYBALL | accuracy: 94.5 -> 95.3 | Helped |
| Use attention-based feature fusion instead of summation (ablation) | Bi-Causal · VOLLEYBALL | accuracy: 94.0 -> 95.8 | Helped |
| Use attention-based feature fusion instead of concatenation (ablation) | Bi-Causal · VOLLEYBALL | accuracy: 94.3 -> 95.8 | Helped |
| Add kinematic features to feature fusion (ablation) | Bi-Causal · VOLLEYBALL | accuracy: 94.7 -> 95.8 | Helped |
| Use all loss functions instead of only the KL-divergence loss (ablation) | Bi-Causal · VOLLEYBALL | accuracy: 91.6 -> 95.8 | Helped |
Limitations
- Authors note the need to improve bi-causality representation.
- Extension to other group activity recognition scenarios is left open.
- Collective Activity lacks fixed interactive objects, so the most salient nearby object is used as a proxy.
Experimental setup
Datasets: VOLLEYBALL RE-ANNOTATED VOLLEYBALL COLLECTIVE ACTIVITY NBA
Benchmarks: VOLLEYBALL RE-ANNOTATED VOLLEYBALL COLLECTIVE ACTIVITY
Metrics: group activity accuracy
Baselines: GroupFormer COMPOSER GIRN ACCG DIN DECOMPL Dual Tamura et al.
Builds on: GroupFormer ARG GIRN mutual learning ViT
Closest papers by meaning
- Detector-Free Weakly Supervised Group Activity Recognition (CVPR 2022 · similarity 0.86)
- A Hierarchical Deep Temporal Model for Group Activity Recognition (CVPR 2016 · similarity 0.85)
- An Actor-Centric Causality Graph for Asynchronous Temporal Inference in Group Activity (CVPR 2023 · similarity 0.84)
- Learning Visual Context for Group Activity Recognition (AAAI 2021 · similarity 0.84)
- Convolutional Relational Machine for Group Activity Recognition (CVPR 2019 · similarity 0.84)
- Generalizable Sensor-Based Activity Recognition via Categorical Concept Invariant Learning (AAAI 2025 · similarity 0.81)
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.