Extracting mAP for each class when evaluating on Custom MS COCO (Class specific mAP evaluation)

[Eldad27]

Hi I have successfully constructed a custom MS COCO dataset (over 1000 images) and attempted to train on yolact. However, there are some issues I'll appreciate some help on.

1. mAP is running far below expectation compared to other instance segmentation algorithms (mask RCNN and RetinaMask) I previously trained on my dataset (mAPs of above 90 for Bbox and above 80 for Masks in all). How do I resolve this?

2. Is there some way to extract a mAP metric for each individual class/label when evaluating? For example, I have an image dataset of let's say, Apples to train. The apples have two labels for 2 classes of 'Good' and 'Bad'. Each Apple/instance can have only one label (Either 'Good' or 'bad').

How do I obtain the separate mAP (and AP if possible) for each class('Good' category and 'bad' category)?

Any help or advice is appreciated, thanks!

Kindly find below the results I'm obtaining after some iterations and my config specification from config.py. I do not want only the overall mAP. I need mAP for each class as well.

Your help is greatly appreciated.

[marcoluis97]

Hello, I am also looking to figure out the same issues as yourself. BY now I still haven't solved any issue, but I'm curious as to why your ETA is so high compared to the training iteration you are in. Could you tell me what GPU are you using?