Hijab Person Rfdetr

--- license: other library_name: rfdetr pipeline_tag: object-detection tags: - object-detection - rf-detr - person-detection - deepstream - onnx --- # hijab-person-rfdetr RF-DETR (DINOv2 backbone) person detector exported to ONNX, used by the `people` viscom in the vidiomatic CV platform. **Private / internal (nextoar) — do not redistribute.** ## Contract — how to bind this model | | | |---|---| | file | `Hijab_model_2.onnx` | | input | `input` float32 `[1, 3, 560, 560]`, **RGB** | | preprocessing | ImageNet norm after /255 — `net-scale-factor=0.01735207357`, `offsets=123.675;116.28;103.53`, **stretched to square** (no letterbox) | | outputs | `dets` `[1,300,4]` (DETR cx,cy,w,h, normalised) · `labels` `[1,300,2]` (per-query logits) | | scoring | **sigmoid**, top-k, **NMS-free** (set predictor) | | classes | **index 0 = person** (single trained class; index 1 unused). Detections fire on `class_id 0`. | ## Training & caveat Trained with RF-DETR (Roboflow) on a single class `hijab_person` (people wearing a hijab), relabeled to `person` for the people viscom. **Recall on people *not* wearing a hijab is not guaranteed** — swap in a broader person detector for general use. ## DeepStream / TensorRT usage Requires the RF-DETR custom output parser `NvDsInferParseCustomRfDetr` (`cluster-mode=4`, NMS-free). On heavy downscale set `scaling-filter=1` (bilinear) — nvinfer's default nearest-neighbour resize collapses RF-DETR recall (~33% vs raw PyTorch). ## Pull ```python from huggingface_hub import hf_hub_download onnx = hf_hub_download("nxtr-kiranshivaraju/hijab-person-rfdetr", "Hijab_model_2.onnx") ```

Source context: 0 downloads · 0 likes · Pipeline object-detection · Library rfdetr · Repo nxtr-kiranshivaraju/hijab-person-rfdetr