RF-DETR Base

roboflow/rf-detr

Labels

Build

Embedl-optimized

Precision

INT8

Runtime

TensorRT 10.16.2.10 with Embedl optimizations

Input size

560×560

Output

Bounding boxes, class labels and confidence scores

RF-DETR Base is a real-time object detector: for each object in an image it returns a bounding box, a class label and a confidence score.

Methodology

Latency: median GPU time per image, batch 1, timed with CUDA events around each run.
Accuracy: COCO box mAP (IoU 0.50 to 0.95), with mAP@50 and mAP@75, on 5,000 val2017 images, each through the served model's own pre- and post-processing.
Memory: peak memory while running the model.
Baseline TensorRT: Roboflow's original model in FP16 on stock TensorRT, same settings, same device.

Image	Size
Model docker.io/embedl/model-artifacts:rf-detr-base.int8-embedl.agx-orin.695c5a2286d94889	38.0 MiB
Runtime docker.io/embedl/models:runtimes.tensorrt.arm64.2026110212523521	703 MiB

Every build of RF-DETR Base on AGX Orin.

Variant	Precision	Speed	Peak memory	Runtime
Embedl-optimized · TensorRT This page huggingface_int8_embedl_tensorrt	INT8	5.951 ms Latency: Median GPU time	631 MiB	TensorRT 10.16.2.10 with Embedl optimizations runtimes.tensorrt.arm64.2026110212523521
Baseline TensorRT rfdetr_fp16_tensorrt	FP16	6.909 ms Latency: Median GPU time	621 MiB	TensorRT 10.16.2.10 runtimes.tensorrt.arm64.2026110212523521