Joint Attention detection on Held-out (test)
77.6AccuracyProposed Model
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Proposed Modelarchitecture=dual-stream Transformer2026.04 | 77.6 | 79.5 | 74.5 | 76.7 | 86.2 | |
| Conv. Baselinearchitecture=lightweight CNN2026.04 | 62.6 | 64.5 | 56.2 | 58.9 | 68.1 | |
| Best LLMtype=Multimodal LLM2026.04 | 59.5 | 58.6 | 63.5 | 61 | 59.5 |