480x640 yolov7 engine give no detection

[michelc33]

When converting a 640x640 onnx model with trtexec ,detections are fine -> OK
When converting a 480x640 onnx model with trtexec , No detections

There is no error in the log and the execution is find except no detection
CPU preprocessing has the same effect

Any direction to look into ?

(NV jetson Orin)

[justincdavis]

Are you able to give more information about how you are building the engines?
Additionally, can you confirm that the engine input sizes are as you would expect with the inspect command like this:
python3 -m trtutils inspect --engine ENGINE_PATH

[michelc33]

(myenv) michelc@jetson1:~/projects/ju$ python -m trtutils inspect --engine yolov7_480_jn.trt
2025-07-21 09:22:52,703 [INFO] trtutils: Loaded engine size: 73 MiB
2025-07-21 09:22:52,781 [INFO] trtutils: [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +43, now: CPU 0, GPU 113 (MiB)
2025-07-21 09:22:52,793 [INFO] trtutils: Engine Size: 42.77 MB
2025-07-21 09:22:52,793 [INFO] trtutils: Max Batch Size: 1
2025-07-21 09:22:52,793 [INFO] trtutils: Inputs:
2025-07-21 09:22:52,793 [INFO] trtutils: 	images: shape=(1, 3, 480, 640), dtype=DataType.FLOAT, format=TensorFormat.LINEAR
2025-07-21 09:22:52,793 [INFO] trtutils: Outputs:
2025-07-21 09:22:52,793 [INFO] trtutils: 	output: shape=(1, 18900, 85), dtype=DataType.FLOAT, format=TensorFormat.LINEAR
(myenv) michelc@jetson1:~/projects/ju$ python -m trtutils inspect --engine yolov7_jn.trt
2025-07-21 09:23:30,640 [INFO] trtutils: Loaded engine size: 73 MiB
2025-07-21 09:23:30,703 [INFO] trtutils: [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +71, now: CPU 0, GPU 141 (MiB)
2025-07-21 09:23:30,714 [INFO] trtutils: Engine Size: 70.63 MB
2025-07-21 09:23:30,714 [INFO] trtutils: Max Batch Size: 1
2025-07-21 09:23:30,714 [INFO] trtutils: Inputs:
2025-07-21 09:23:30,715 [INFO] trtutils: 	images: shape=(1, 3, 640, 640), dtype=DataType.FLOAT, format=TensorFormat.LINEAR
2025-07-21 09:23:30,715 [INFO] trtutils: Outputs:
2025-07-21 09:23:30,715 [INFO] trtutils: 	num_dets: shape=(1, 1), dtype=DataType.INT32, format=TensorFormat.LINEAR
2025-07-21 09:23:30,715 [INFO] trtutils: 	det_boxes: shape=(1, 100, 4), dtype=DataType.FLOAT, format=TensorFormat.LINEAR
2025-07-21 09:23:30,715 [INFO] trtutils: 	det_scores: shape=(1, 100), dtype=DataType.FLOAT, format=TensorFormat.LINEAR
2025-07-21 09:23:30,715 [INFO] trtutils: 	det_classes: shape=(1, 100), dtype=DataType.INT32, format=TensorFormat.LINEAR

The yolov7_jn is giving me correct detection
It looks like the output tensort has a different format, NMS related ?
Thks

[justincdavis]

It would be a NMS/output postprocessing issue.

Currently, the supported output shapes I have implemented are (1, 300, 6) for YOLOv10 style outputs, and then the 4 output arrays from EfficientNMS TRT plugin as seen in the second model you inspect.

If you export your model with the --end2end flag from the YOLOv7 repository you should get the correct NMS output tagged.

You can follow this page on what arguments you should supply for YOLOv7 export.
https://trtutils.readthedocs.io/en/latest/tutorials/yolo/yolov7.html