levipereira / triton-client-yolo Goto Github PK

Hello,

First of all, thanks for this repo.

Could I ask you why in the resulting file, _predictions.json, the number of objects per image is so high?

The triton server is launched with the NMS plugin, so I'm a bit confused, to be honest, to see so many detections per image.

Thank you

Hello, while using the code in this repository I noticed an issue with the width and height parameters that impacts rectangular-shaped models:

1 - In client.py the shape is being passed as NCWH, rather than NCHW, which doesn't seem correct.

2- In preprocess the code clearly expects input_shape to be [height, width], while is being passed as [width, height] from client.py

3- In postprocess input_shape is used in a way that it's consistent to [width, height], based on the coordinates passed to BoungingBox ctor and the scaling. This is consistent with how they are being passed, but inconsistent with the preprocess method.