Articles
Guides and comparisons for people who deploy detection models. Every number is sourced from a verified run, and the charts are live.
- July 3, 2026 · 3 min read
Best object detection models for Jetson Orin Nano (2026)
On the Jetson Orin Nano Super with TensorRT FP16, D-FINE-X leads accuracy at 61.4 mAP and DEIMv2-Atto runs fastest at 89.1 FPS.
Read - July 3, 2026 · 3 min read
Best object detection models for RTX 5070 Ti (2026)
On the RTX 5070 Ti with TensorRT FP16, D-FINE-X leads at 61.5 mAP and RT-DETRv2-R34 runs fastest at 132.9 FPS. Accuracy is the deciding axis.
Read - July 3, 2026 · 3 min read
Best object detection models for Pi 5 + Hailo-8 (2026)
On a Raspberry Pi 5 with the Hailo-8 NPU at INT8, YOLOv9-C leads accuracy at 54.8 mAP and YOLOX-Tiny runs fastest at 195.6 FPS. Pick by latency budget.
Read - July 3, 2026 · 3 min read
Best object detection models for Raspberry Pi 5 (2026)
On Raspberry Pi 5 with ONNX Runtime, D-FINE-X tops accuracy at 61.4 mAP while DEIMv2-Atto runs fastest at 30.3 FPS. Pick by latency budget.
Read - July 3, 2026 · 4 min read
DEIM-L vs RF-DETR-L: benchmark on COCO
RF-DETR-L leads by 0.8 mAP and is 25.84% faster in PyTorch, but DEIM-L wins on ONNX, TensorRT, Jetson, and mAP per GFLOP.
Read - July 3, 2026 · 4 min read
DEIM-L vs RT-DETRv2-R101: same-protocol benchmark
Measured on the same COCO protocol: DEIM-L matches RT-DETRv2-R101 on accuracy with 40% of the parameters and half the VRAM, and overtakes it on TensorRT.
Read - July 3, 2026 · 4 min read
DEIM-L vs RT-DETRv2-R50: benchmark on same protocol
Measured on the same COCO protocol: DEIM-L leads by 2.1 mAP points with less VRAM, but RT-DETRv2-R50 is 42% faster in PyTorch on an RTX 5070 Ti.
Read - July 3, 2026 · 3 min read
DEIM-L vs YOLOX-L: benchmark on one protocol
Measured on the same COCO protocol: DEIM-L leads by 2.4 mAP points, but YOLOX-L is 56% faster in desktop PyTorch and the order flips on TensorRT.
Read - July 3, 2026 · 4 min read
DEIM-N vs YOLOv9-T: benchmark on the same protocol
Measured on the same COCO protocol: DEIM-N leads YOLOv9-T by 5.0 mAP points and is 15.84% faster on desktop GPU, but the order flips on Pi 5.
Read - July 3, 2026 · 4 min read
DEIM-S vs YOLOv9-S: benchmark on the same protocol
Measured on the same COCO protocol: DEIM-S leads YOLOv9-S by 1.7 mAP points and is 8.42% faster on desktop GPU, but YOLOv9-S wins on Pi 5.
Read - July 3, 2026 · 4 min read
DEIM vs DEIMv2: full benchmark comparison
DEIMv2 wins 4 of 5 matched pairs on accuracy, but DEIM runs 10.4% faster on average on an RTX 5070 Ti.
Read - July 3, 2026 · 4 min read
DEIM vs D-FINE: which detector family wins
D-FINE wins 4 of 5 matched pairs and reaches 61.4 mAP, edging DEIM on accuracy at nearly equal speed on an RTX 5070 Ti.
Read - July 3, 2026 · 4 min read
DEIM vs RT-DETR: which detector family wins
Measured on one COCO protocol on an RTX 5070 Ti: DEIM averages 4.3 mAP points higher at matched compute, while RT-DETR runs 41.6% faster.
Read - July 3, 2026 · 4 min read
D-FINE-L vs YOLOX-L: benchmark on the same protocol
Measured on the same COCO protocol: D-FINE-L leads by 4.6 mAP points with less VRAM, but YOLOX-L runs 49% faster in PyTorch on an RTX 5070 Ti.
Read - July 3, 2026 · 4 min read
D-FINE vs RF-DETR: full benchmark comparison
Measured on one COCO protocol on an RTX 5070 Ti: D-FINE averages 7.5 mAP points higher at matched compute, while RF-DETR runs 35.8% faster.
Read - July 3, 2026 · 4 min read
D-FINE vs RT-DETR: which detector family wins
D-FINE averages 6.4 mAP points more than RT-DETR at matched compute, while RT-DETR runs 33.8% faster on an RTX 5070 Ti.
Read - July 3, 2026 · 4 min read
D-FINE vs RT-DETRv2: full benchmark comparison
Measured on one COCO protocol on an RTX 5070 Ti: D-FINE averages 6.2 mAP points higher at matched compute, while RT-DETRv2 runs 39.7% faster.
Read - July 3, 2026 · 4 min read
D-FINE vs YOLOv9: which detector family wins
Measured on one COCO protocol on an RTX 5070 Ti: D-FINE averages 2.5 mAP points higher at matched compute, while YOLOv9 runs 19.9% faster.
Read - July 3, 2026 · 3 min read
ncnn FP32 vs ONNX FP32 on Raspberry Pi 5: 13 models
On Raspberry Pi 5 CPU, ncnn FP32 runs a median 2.09x faster than ONNX Runtime FP32 across 13 models, but the YOLOX models give up a point of mAP.
Read - July 3, 2026 · 3 min read
ONNX FP32 vs PyTorch FP32 on Jetson Orin: 55 models
ONNX Runtime FP32 gives only a median 1.02x over PyTorch FP32 across 55 models on Jetson Orin, and many models run slower. TensorRT is the better path.
Read - July 3, 2026 · 3 min read
ONNX FP32 vs PyTorch FP32 on RTX 5070 Ti: 55 models
ONNX Runtime FP32 gives a median 1.92x speedup over PyTorch FP32 across 55 models on the RTX 5070 Ti, with no accuracy loss and only one model slower.
Read - July 3, 2026 · 3 min read
ONNX FP32 vs PyTorch FP32 on Raspberry Pi 5: 48 models
ONNX Runtime FP32 gives a median 1.57x speedup over PyTorch FP32 across 48 models on Raspberry Pi 5, but it costs accuracy and one model runs slower.
Read - July 3, 2026 · 4 min read
Commercially safe object detection models for A100
YOLO-NAS ships under a non-permissive license. On the A100, RT-DETR-R50m (Apache-2.0) matches YOLO-NAS-L at 50.8 vs 51.2 mAP.
Read - July 3, 2026 · 4 min read
Commercially safe object detection models for Jetson Orin
YOLO-NAS ships under a non-permissive license. On Jetson Orin, YOLOv9-M matches it at 56.1 mAP under a permissive MIT license.
Read - July 3, 2026 · 4 min read
Commercially safe detection models for Pi 5 + Hailo-8
YOLO-NAS ships under a non-permissive license. On a Pi 5 with Hailo-8, YOLOv9-M (MIT) beats YOLO-NAS-L by 0.9 mAP at INT8.
Read - July 3, 2026 · 4 min read
Commercially safe object detection models for Raspberry Pi 5
YOLO-NAS ships under a non-permissive license. On the Raspberry Pi 5, RF-DETR-S (Apache-2.0) matches YOLO-NAS-M at 55.1 vs 55.4 mAP and runs faster.
Read - July 3, 2026 · 4 min read
PicoDet vs YOLOX: which detector family wins
Measured on one COCO protocol on an RTX 5070 Ti: PicoDet averages 5.1 mAP points higher in the sub-5M tier, while YOLOX runs 21.3% faster and scales far larger.
Read - July 3, 2026 · 4 min read
RF-DETR-N vs RT-DETRv2-R18: same-protocol benchmark
Measured on the same COCO protocol: RF-DETR-N edges RT-DETRv2-R18 by 0.7 mAP at 384 px, but RT-DETRv2-R18 runs at 640 px and is faster on GPU.
Read - July 3, 2026 · 4 min read
RF-DETR vs RT-DETR: which detector family wins
Measured on one COCO protocol on an RTX 5070 Ti: RF-DETR averages 1.6 mAP points higher at matched compute, while RT-DETR runs about 3% faster.
Read - July 3, 2026 · 4 min read
RF-DETR vs YOLOv9: where the frontier crosses
The frontier crosses: YOLOv9 wins the small end and runs 22.8% faster, while RF-DETR tops out at 58.6 mAP on an RTX 5070 Ti.
Read - July 3, 2026 · 4 min read
RF-DETR vs YOLOX: full benchmark comparison
Measured on one COCO protocol on an RTX 5070 Ti: RF-DETR averages 3.9 mAP points higher at matched compute, while YOLOX runs 32.2% faster.
Read - July 3, 2026 · 4 min read
RT-DETR vs RT-DETRv2: full benchmark comparison
RT-DETR and RT-DETRv2 tie on accuracy within 0.1 mAP points, but RT-DETRv2 runs 11.7% faster on an RTX 5070 Ti.
Read - July 3, 2026 · 4 min read
RT-DETR vs YOLOv9: full benchmark comparison
YOLOv9 beats RT-DETR at matched compute, averaging 4.0 mAP points higher and running 7.2% faster on an RTX 5070 Ti.
Read - July 3, 2026 · 4 min read
RT-DETR vs YOLOX: which detector family wins
RT-DETR wins 6 of 7 matched pairs on accuracy, but YOLOX runs 30.3% faster on average on an RTX 5070 Ti.
Read - July 3, 2026 · 3 min read
TensorRT FP16 slower than FP32 on the RTX 5070 Ti
On an RTX 5070 Ti, TensorRT FP16 measured slower than FP32 for many models: D-FINE-N ran 79.7 FPS against 98.7 on FP32.
Read - July 3, 2026 · 3 min read
TensorRT FP16 vs PyTorch FP32 on Jetson Orin: 55 models
TensorRT FP16 gives a median 3.39x speedup over PyTorch FP32 across 55 models on Jetson Orin, and only one model loses more than half a point of mAP.
Read - July 3, 2026 · 3 min read
TensorRT FP16 vs PyTorch FP32 on RTX 5070 Ti: 51 models
TensorRT FP16 gives a median 2.22x speedup over PyTorch FP32 across 51 models on the RTX 5070 Ti, but four models lose accuracy and one runs slower.
Read - July 3, 2026 · 3 min read
TensorRT FP32 vs PyTorch FP32 on Jetson Orin: 55 models
TensorRT FP32 gives a median 2.2x speedup over PyTorch FP32 across 55 models on Jetson Orin, with no accuracy loss and no model running slower.
Read - July 3, 2026 · 4 min read
YOLOv9 vs YOLOX: which detector family wins
At matched compute YOLOv9 averages 5.3 mAP points more than YOLOX, while YOLOX runs 24.6% faster on an RTX 5070 Ti.
Read
In progress
Drafts on the way. They publish once every number in them comes from a verified run.
- comparison · edge · yolov9In progress
YOLOv9-S vs YOLOX-S: choosing a small detector
Two detectors near 9M parameters, measured on the same COCO protocol. Where each one wins on accuracy, latency, and edge support.
- comparison · deim · rtdetrIn progress
DEIM-L vs RT-DETR-R101: benchmark on COCO
DEIM-L leads by 1.0 mAP on under half the parameters, and turns RT-DETR-R101's PyTorch speed lead around once you convert the model.
- comparison · deim · rtdetrIn progress
DEIM-L vs RT-DETR-R50: benchmark on the same protocol
Measured on the same COCO protocol: DEIM-L leads by 1.9 mAP points with fewer parameters, but RT-DETR-R50 is 37.5% faster on desktop GPU.
- comparison · deim · rtdetrIn progress
DEIM-L vs RT-DETR-R50m: benchmark on one protocol
Measured on the same COCO protocol: DEIM-L leads by 4.0 mAP points, but RT-DETR-R50m is 47% faster on desktop GPU and the speed lead vanishes on a Raspberry Pi 5.
- comparison · deim · rtdetrIn progress
DEIM-L vs RT-DETR-X: benchmark on one protocol
Measured on the same COCO protocol: accuracy ties at 57.8 vs 57.9 mAP, but DEIM-L reaches it with half the parameters and 245 vs 401 MB peak VRAM.
- comparison · deim · rtdetrIn progress
DEIM-M vs RT-DETR-R34: benchmark on the same protocol
Measured on the same COCO protocol: DEIM-M leads by 3.3 mAP points with fewer parameters, but RT-DETR-R34 is 29.41% faster on desktop GPU.
- comparison · deim · deimv2In progress
DEIM-S vs DEIMv2-S: measured on the same protocol
DEIMv2-S measures 0.9 mAP points higher than DEIM-S and uses less VRAM, but DEIM-S runs 37% faster on the same COCO protocol.
- comparison · deim · dfineIn progress
DEIM-S vs D-FINE-S: measured on the same protocol
DEIM-S and D-FINE-S share 10.32M parameters; on the same COCO protocol D-FINE-S measures 1.3 mAP points higher on desktop GPU.
- comparison · deim · rtdetrIn progress
DEIM-S vs RT-DETR-R18: benchmark on the same protocol
Measured on the same COCO protocol: DEIM-S leads by 2.3 mAP points with fewer parameters, but RT-DETR-R18 is 29.37% faster on desktop GPU.
- comparison · dfine · rfdetrIn progress
D-FINE-L vs RF-DETR-L: benchmark on the same protocol
Measured on the same COCO protocol: D-FINE-L leads by 1.4 mAP points and is faster on every converted runtime, though the two run at different input sizes.
- comparison · dfine · rtdetrIn progress
D-FINE-L vs RT-DETR-R101: same-protocol benchmark
Measured on the same COCO protocol: D-FINE-L leads by 3.2 mAP points with fewer than half the parameters, and beats RT-DETR-R101 on most runtimes.
- comparison · dfine · rtdetrIn progress
D-FINE-L vs RT-DETR-R50: benchmark on the same protocol
Measured on the same COCO protocol: D-FINE-L leads by 4.1 mAP points with fewer parameters, but RT-DETR-R50 is 26.58% faster on desktop GPU.
- comparison · dfine · rtdetrIn progress
D-FINE-L vs RT-DETR-R50m: benchmark on protocol
Measured on the same COCO protocol: D-FINE-L leads by 6.2 mAP points, but RT-DETR-R50m is 38% faster on desktop GPU and that lead vanishes on a Raspberry Pi 5.
- comparison · dfine · rtdetrIn progress
D-FINE-M vs RT-DETR-R34: benchmark on one protocol
Measured on the same COCO protocol: D-FINE-M leads by 5.6 mAP points, but RT-DETR-R34 is 28% faster on desktop GPU and that lead vanishes on a Raspberry Pi 5.
- comparison · dfine · rtdetrIn progress
D-FINE-M vs RT-DETR-R50m: same-protocol benchmark
Measured on the same COCO protocol: D-FINE-M leads by 4.0 mAP points on half the VRAM, but RT-DETR-R50m is faster on desktop GPU and the order flips on edge.
- comparison · dfine · yolov9In progress
D-FINE-N vs YOLOv9-T: benchmark on the same protocol
Measured on the same COCO protocol: D-FINE-N leads by 4.0 mAP points, but YOLOv9-T is lighter, more compute-efficient, and faster on Raspberry Pi 5.
- comparison · dfine · rtdetrIn progress
D-FINE-S vs RT-DETR-R18: benchmark on one protocol
Measured on the same COCO protocol: D-FINE-S leads by 3.6 mAP points, but RT-DETR-R18 is 27% faster on desktop GPU and the order flips on a Raspberry Pi 5.
- comparison · dfine · yolov9In progress
D-FINE-S vs YOLOv9-S: benchmark on the same protocol
Measured on the same COCO protocol: D-FINE-S leads by 3.0 mAP points on desktop GPU, but YOLOv9-S is the faster detector on Raspberry Pi 5.
- comparison · rfdetr · rtdetrIn progress
RF-DETR-N vs RT-DETR-R18: benchmark on COCO
RF-DETR-N leads by 1.6 mAP points at 384 px, but RT-DETR-R18 at 640 px is faster on desktop and every GPU runtime measured.
- comparison · rfdetr · yolov9In progress
RF-DETR-N vs YOLOv9-C: benchmark on COCO
YOLOv9-C leads by 5.7 mAP points at 640 px, but RF-DETR-N at 384 px pulls ahead on Jetson Orin and Raspberry Pi 5.
- comparison · rfdetr · yoloxIn progress
RF-DETR-N vs YOLOX-L: benchmark on COCO
YOLOX-L leads by 4.0 mAP points at 640 px, but RF-DETR-N at 384 px is faster on edge hardware and most converted runtimes.
- comparison · rfdetr · yoloxIn progress
RF-DETR-N vs YOLOX-M: benchmark on one protocol
Measured on the same COCO protocol at each model's default input: accuracy ties at 51.4 vs 51.7 mAP, and which one is faster depends entirely on the runtime.
- comparison · rfdetr · yolov9In progress
RF-DETR-S vs YOLOv9-C: benchmark on COCO
YOLOv9-C leads by 2.0 mAP points at 640 px, but RF-DETR-S at 512 px is faster on ONNX, Jetson Orin, and Raspberry Pi 5.
- comparison · rfdetr · yoloxIn progress
RF-DETR-S vs YOLOX-L: benchmark on one protocol
Measured on the same COCO protocol at each model's default input: accuracy ties at 55.1 vs 55.4 mAP, and RF-DETR-S uses less than a third the peak VRAM.
- comparison · rtdetr · rtdetrv2In progress
RT-DETR-R18 vs RT-DETRv2-R18: benchmark
On the same COCO protocol, RT-DETRv2-R18 measures 0.9 mAP points higher than RT-DETR-R18 and runs 51.1 FPS to 47.2 on desktop GPU.
- comparison · rtdetr · yolov9In progress
RT-DETR-R18 vs YOLOv9-C: measured on the same protocol
On the same COCO protocol, YOLOv9-C leads by 7.3 mAP points while RT-DETR-R18 runs 47.2 FPS to 42.3 on desktop GPU.
- comparison · rtdetr · yolov9In progress
RT-DETR-R34 vs YOLOv9-C: benchmark on one protocol
On the same COCO protocol YOLOv9-C leads RT-DETR-R34 by 4.9 mAP points, yet RT-DETR-R34 is faster on every runtime except PyTorch.
- comparison · rtdetr · yoloxIn progress
RT-DETR-R34 vs YOLOX-L: benchmark on one protocol
On the same COCO protocol YOLOX-L leads RT-DETR-R34 by 3.2 mAP points, yet RT-DETR-R34 is faster on every runtime except PyTorch.
- guide · rtmdet · libreyoloIn progress
How to run RTMDet with LibreYOLO
A step-by-step guide to loading and benchmarking RTMDet through the LibreYOLO API.
- guide · edge · jetsonIn progress
Picking an edge detector for Jetson Orin Nano
Reading the accuracy-vs-latency frontier for embedded GPU deployment, with TensorRT in the mix.