Face_Direction

#Work

Archived at 9/15/2026, 3:38:51 PM

• [-] Retrain and Validate Person Attributes Model 🆔 bp0t9x ⛔ isp2fp Retrain the multi-task/attribute recognition model with the newly balanced dataset to boost recall on the target attributes.
  • Run training with the updated data distribution.
  • Ensure performance on existing major classes (like vehicle color or type) hasn't degraded (no catastrophic forgetting).

Archived at 9/15/2026, 3:38:33 PM

• [-] Retrain and Validate Person Attributes Model 🆔 bp0t9x ⛔ isp2fp Retrain the multi-task/attribute recognition model with the newly balanced dataset to boost recall on the target attributes.
  • Run training with the updated data distribution.
  • Ensure performance on existing major classes (like vehicle color or type) hasn't degraded (no catastrophic forgetting).

Archived at 9/15/2026, 3:38:02 PM

• [-] Retrain and Validate Person Attributes Model 🆔 bp0t9x ⛔ isp2fp Retrain the multi-task/attribute recognition model with the newly balanced dataset to boost recall on the target attributes.
  • Run training with the updated data distribution.
  • Ensure performance on existing major classes (like vehicle color or type) hasn't degraded (no catastrophic forgetting).

Archived at 9/15/2026, 3:36:03 PM

• [-] Data Augmentation & Balancing for "Accessories" and "Suits" 🆔 isp2fp Address the class imbalance for person attributes. Gather, filter, or synthetically augment data to increase the representation of "accessories" and "suits" (apparel).

  • Run vlm on the person crops to increase dataset size.
  • Query the existing dataset to check the exact distribution of "accessory" and "suit" tags.
  • Apply targeted data augmentation (e.g., color jittering, cropping, or mixing) specifically on underrepresented classes.
  • (Optional) Source or scrape additional data points if augmentation isn't enough.

• implement face direction using facial landmarks on triton.

  • convert the github repo to work with triton.
  • detect if a face is facing forward.
  • add it to the face ensemble.
  • neglect faces not facing forward.