OpenVeil-Verifier-DINOv3-ViT-L16
verifying people in video frames for anonymization
2 min readupdated
This model checks whether a cropped video frame actually contains a person so it can automatically pixelate them before a video dataset is published. Developed by researchers at Cornell University, it is a binary classifier built on facebook/dinov3-vitl16-pretrain-lvd1689m (DINOv3 ViT-L/16) with a LoRA (a small adapter layered over the base model) and an MLP head. Unlike standard detectors, it acts as a second stage after Grounding DINO, filtering out false positives at a low confidence threshold so no person slips through unblurred. It is useful for researchers and dataset builders who need to quickly mask people in surveillance footage.
Key features
- High filtering accuracy: achieves a cross-validation ROC AUC of 0.994, and on the HUM731 benchmark, it pixelated 172 out of 172 people (covering at least 90% of their area).
- Fast inference: on a single NVIDIA A10 GPU in
bfloat16with merged weights, it processes 276.7 crops per second (or 191.8 per second including CPU frame preprocessing). - Resistant to false positives: trained on 50,756 farm video crops where the main distractions were calves, cows, and pen structures.
- Safe weight format: the adapter and classification head are distributed as
safetensorswith checksums, avoiding unsafe pickle files.
What you can do
- Plug it into the OpenVeil pipeline to automatically blur staff and visitors in scientific or industrial footage before releasing it publicly.
- Use it as a starting point for fine-tuning a person verifier tailored to your own cameras, warehouses, or other environments.
- Pair it with Grounding DINO as a secondary filter when your primary detector triggers too many false positives on animals or background clutter.
Example
from openveil.detect.verifier import CropVerifier
verifier = CropVerifier(
adapter_dir="weights/OpenVeil-Verifier-DINOv3-ViT-L16",
device="cuda"
).load()
scores = verifier.score_frames([frame_bgr], [[[x1, y1, x2, y2]]])
Read more
DetailsMetrics: ROC AUC 0.994; across 2,000 everyday frames, false boxes appear in 3.3% of cases (large ones ≥15% of the frame in 0.4%).
Architecture: LoRA (rank 32, alpha 64) on all linear layers of DINOv3 ViT-L/16 + MLP head 1024 → 1024 → 512 → 2, 224×224 input.
Requirements and compatibility: requires
transformers ≥ 4.56 and < 5 (tested on 4.57.1 + peft 0.13.2); access to the base DINOv3 model must be requested separately from Meta.Limitations: trained solely on calf barn footage; performance on night-vision IR footage with people has not been tested, and partially occluded legs or people near frame edges may be missed.










Comments