Does anyone knows the preprocessing and postprocessing of YOLOv2 coco from onnx zoo?

[Apisteftos]

I found a code about the tiny yolov2 onnx and followed the instructions from C# and darknet about the anchors and preprocessing.

I am using the

anchors = [0.57273, 0.677385, 1.87446, 2.06253, 3.33843, 5.47434, 7.88282, 3.52778, 9.77052, 9.16828]

tooked from darknet. Now the postprocessing should be similar to tiny yolov2 which seems isn't.

I am getting wrong class prediction + the bounding boxes are not in place.

[xadupre]

Is it possible to know more about the models you are comparing? (location, converter you used, ...)

[Apisteftos]

Is it possible to know more about the models you are comparing? (location, converter you used, ...)

I am not converting them, I am using these from the onnx zoo. I am doing my thesis in AI accelerators and testing the performance of the pre-trained onnx models from onnx zoo. I did that for almost all classification models and couple of them of the object detections & image segmentation, but they have huge problems.

Problems:

• Incomplete information ( preprocessing, postprocessing alogrithms, mAP from object detection model does not corresponds to the real numbers. Sometime is higher or less than what is documented )
• GPU acceleration (some convolution nodes exceeding the theoretical peak TFLOPs during the inference, which is good, but I don't know how to estimate it and maybe no one do it. )
• Profiling (missing data to estimate the performance , i.e. qlinearconv, MaxPool the are not extracting the weights in the profiled data, which I need them.)

Tested from Object Detection:

• Yolo v4
• Yolo v3
• Tiny Yolo V3
• Tiny Yolo V2
• SSD
• Faster R-CNN and Mask R-CNN ( which both I should take them out because they had very high performance in 2 nodes. For example 480 node and node 490 convolution operator have for batch size 1, 600TFLOs and 1.3 PFLOPs )
• Yolo V2 attempted but with no results (missing information of preprocessing and mostly for postprocessing algorithms)