You can deploy this model and start earning money today!

MKSHRESTHA/
rtdetr-l_carparts-seg

Name: rtdetr-l_carparts-seg
Availability: InStock

onnx

Views

No views yet

RT-DETR-L Car-Parts Inference (ONNX)

Run object detection on images/videos with the exported best.onnx model (RT-DETR-L fine-tuned on 23 car-part classes).

Model

Architecture: RT-DETR-L
Task: detection (23 classes)
Classes: back_bumper, back_door, back_glass, back_left_door, back_left_light, back_light, back_right_door, back_right_light, front_bumper, front_door, front_glass, front_left_door, front_left_light, front_light, front_right_door, front_right_light, hood, left_mirror, object, right_mirror, tailgate, trunk, wheel
Input: 640x640 (fixed in the ONNX graph)
File: weights/best.onnx

Install

pip install ultralytics onnxruntime opencv-python

(onnxruntime is the CPU build; use onnxruntime-gpu for CUDA.)

Usage

bash

python infer.py --source input.jpg      # single image
python infer.py --source input.mp4      # video
python infer.py --source ./folder       # folder of images
python infer.py --source 0 --show       # webcam, live window

Flags

Flag	Default	Notes
`--source`	required	image / video / folder / webcam index
`--conf`	`0.1`	confidence threshold
`--imgsz`	`640`	input size
`--device`	`""`	`""` = ONNXRuntime auto (CPU/GPU), or `"0"` for a GPU
`--nosave`	off	skip writing output files
`--show`	off	show a live window

Output

Annotated results are saved to:

weights/runs/detect/predict/        (first run)
weights/runs/detect/predict2/       (next run, etc.)

Images: copy with bounding boxes drawn.
Video: new .mp4 with boxes on every frame (input is never overwritten).

Important: confidence threshold

The exported model scores every detection below 0.25 (typical max ≈ 0.16). This is normal DETR score calibration and does not mean the model is bad — mAP is based on score ranking, not absolute value. Because of this, the default --conf is set to 0.1. If you see no boxes, lower it:

python infer.py --source input.jpg --conf 0.05

If you get too many overlapping/duplicate boxes, tighten NMS:

python infer.py --source input.jpg --conf 0.1 --iou 0.5

Notes / troubleshooting

Use python3 if python is not on your PATH.
The ONNX graph is fixed at 640x640; other --imgsz values are letterboxed.
To use the PyTorch weights instead, point the script at best.pt (MODEL = "best.pt" at the top of infer.py).

MKSHRESTHA/rtdetr-l_carparts-seg