v1 — Initial release

v1 — Initial release

Implementation code and resources for RAFNet: A Relation-Aware Adaptive Fusion Network for Dense Classroom Student Behavior Recognition (Xia et al., 2026).

Assets

rafnet_features.tar.gz — 726 MB

Pre-extracted GroupRec + ConvNeXt features on NCST Classroom, with bbox_cxcywh_norm / img_uid context fields required by C+Spatial.

• Contents: prism_features/latest/{train,val,test}_features_with_context.pkl
• 1024-D pooled embeddings; not face-recoverable.
• Extract: tar xf rafnet_features.tar.gz -C output/

SHA256: 46f6f8b082bde00996ceae8461a48b5e84b5619df54e365a1db005ba2612bbb9

rafnet_checkpoints.tar.gz — 179 MB

Trained fusion head checkpoints for 15 paper-reported configs × 5 seeds = 75 checkpoints.

Per seed: best_model.pth (by val Macro F1 no-null), results.json (test metrics + provenance), config.yaml (frozen snapshot).

Models included (15):

• Single-modality: convnext_only, convnext_only_loss_tuned, grouprec_only_fair_ncst, grouprec3d_only, resnet_only_loss_tuned, vit_only_loss_tuned
• MLP fusion: grouprec_convnext_mlp, grouprec_resnet_loss_tuned, grouprec_vit_loss_tuned
• Gated / attention: grouprec_convnext_gated, grouprec_convnext_gated_loss_tuned, grouprec_convnext_cross_attention_loss_tuned
• C+Spatial family: grouprec_convnext_gated_c_spatial_graph_loss_tuned (SOTA), grouprec3d_convnext_gated_loss_tuned, grouprec3d_convnext_gated_c_spatial_graph_loss_tuned

Extract: tar xf rafnet_checkpoints.tar.gz (creates results/training/...).

SHA256: 35b4746a6397a8fe8155197f39de677d88392839ffffda96c348022c3e95348d

Reproduce

pixi install
pixi run eval   # aggregates 75 checkpoints, writes results/evaluation/final_results.json

See the README for full Quick Run and End-to-End paths.