Embedl's build embedl/rf-detr-base of the original model roboflow/rf-detr
Baseline
- embedl/rf-detr-base
- FP16
- AGX Thor
- Labels
- Image
- Detection
- Build
- Baseline
- Precision
- FP16
- Runtime
- TensorRT 10.16.2.10
- Input size
- 560×560
- Classes
- 80 COCO categories
RF-DETR Base (Roboflow) takes a photo and answers with the objects in it: each of the 80 COCO categories it finds, named, scored and boxed.
RF-DETR is Roboflow’s, under the Apache License 2.0 (LICENSE-RF-DETR beside the files).
Methodology
- Latency: on the NVIDIA Jetson AGX Thor, batch 1. The engine: GPU time of one execution between CUDA events, copies excluded, 8622 runs. Per image: wall time of serve.py's stages (decode, resize and normalize, engine call with copies, post-processing) over the 5000 evaluation images.
- Accuracy: COCO val2017 (5000 images, instances_val2017.json), pycocotools' bbox mAP@[.50:.95] (COCOeval, maxDets 100) of the top 300 (query, class) pairs per image, through the code that is served. The square 560×560 resize is what RFDETR.predict() does; upstream's FP32 model scores 53.3 under it.
- Size: the serialized TensorRT engine on disk.
- PyTorch FP32: upstream’s FP32 model under the same square 560×560 resize, as the accuracy reference.