DEIMv2.axera

DEIMv2 Real-Time Object Detection Meets DINOv3, running on Axera NPU (AX650).

This repo contains the deployment-ready ONNX graph, the Pulsar2 build configuration, calibration data, and a minimal inference script.

Model	Note
deimv2_pico.onnx	DEIMv2-HGNetv2-Pico, 640x640, with in-graph post-processing.
deimv2_s.onnx	DEIMv2-DINOv3-S, 640x640, with in-graph post-processing.
deimv2_x.onnx	DEIMv2-DINOv3-X, 640x640, with in-graph post-processing

Results

COCO val2017 (all 5,000 images), bbox mAP@0.50:0.95 using COCOeval and a score threshold of 0.001. The axmodels were evaluated on an AX650A board in NPU1 mode.

Model	Official FP32 ONNX	AX650A axmodel	NPU1 Latency(ms)	NPU3 Latency(ms)
Pico	0.384	0.375	~10.193	5.389
S	0.506	0.506	~155	42.395
X	0.567	0.577	~355	~118

Build

Requires the Axera Pulsar2 toolchain (pulsar2).

pulsar2 build --config config.json --input onnx/deimv2_s.onnx

Inference on board

Copy compiled.axmodel, infer.py and a test image to the board, then:

python3 infer.py -m compiled.axmodel -i test.jpg

Or

python3 infer.py -m compiled.axmodel -d /path/to/images --thr 0.6 -o vis/

Dependencies on board: axengine, numpy, Pillow.

On PC (FP32 reference):

python3 infer.py -m onnx/deimv2_s.onnx -i test.jpg

Note: freezing orig_target_sizes for production

orig_target_sizes only tells the post-processor how to scale boxes back to the original image size. If your deployment always feeds images of one fixed resolution (e.g. a 1920x1080 camera stream), you can hard-code it and drop the second input: in the ONNX graph, replace the orig_target_sizes graph input with an initializer holding [w, h], remove the matching input_configs entry from config.json, and rebuild. Do not do this if input resolutions vary — boxes would be scaled with the wrong size.

License

BSD 3-Clause License. See LICENSE.