Yolact: Performance Improvement on MSCOCO

[abhigoku10]

@dbolya how to improve the accuracy of detection for MSCOCO i.e increase the mAP value , since we have got improved performance

[dbolya]

Umm, can you clarify?

[abhigoku10]

i want to increase the mAP values for the person class and car class for MSCOCO dataset

[dbolya]

Do you want to keep the same speed? Because if I knew of a way to increase performance while keeping speed the same, I'd have done it by now.

I guess you can add more data for those classes from other datasets, but that's not increasing mAP on MSCOCO anymore.

[abhigoku10]

sure i will try adding more classes and observe the results on this whether it will improve or not

[abhigoku10]

@dbolya thanks for the wonderful explanation , i wanted to get into further details . The detection of person from the model for a score threshold of 0.25 has less detection but if i decrease the detection to 0.12 i get more detection along with false positive . how understand from the protonet , mask coefficients how the person detection is varying based on the scenarios like ped walking on the cross walk or ped standing in the long distance #171

[dbolya]

The score is just calculated like any other object detector. It's just predicted along with the mask coefficients for each class, and then we take the score for the highest class.

It's standard to get less false negatives and more false positives as you tune the threshold, which is the entire reason we use mAP and not some accuracy metric (mAP is score threshold agnostic). You just need to find a good threshold for your application, which may be between 0.12 and 0.25. I've been using 0.15 with good success when I've tested it.