Activity Recognition

Activity Recognition

The real-data component study evaluates local activity evidence using MEVA clips.

Clip construction

• 16 frames per activity clip;
• 112 × 112 spatial resolution;
• 28 activity classes;
• actor-centric mode crops around the annotated actor tube before resizing.

Important

The actor tube is annotation-derived. This is an oracle-localization experiment that isolates recognition after localization. It must not be described as an end-to-end detector-plus-recognizer result.

Backbone

The evaluated classifier uses R3D-18 initialized from video pretraining and followed by a 28-class classification head.

Training configuration represented in the research record

• optimizer: AdamW;
• learning rate: 1e-4;
• weight decay: 1e-4;
• label smoothing: 0.05;
• gradient clipping: 1;
• model selection: validation macro-F1;
• ReduceLROnPlateau scheduling;
• early stopping after three epochs without improvement;
• seeds: 1, 21, 42;
• batch size: 1 in the reported environment.

Class-imbalance strategies

Four principal families are compared:

1. no explicit balancing;
2. tempered inverse-frequency weighted loss;
3. weighted random sampling;
4. weighted loss + sampling.

The unweighted actor-centric fully fine-tuned family has the best mean macro-F1 across the three reported seeds, while sampling-based strategies show larger seed sensitivity.

Calibration

Temperature scaling is fitted on validation logits and changes confidence values rather than class labels. ECE and multiclass Brier score are reported alongside discrimination metrics.