Recreated the model, but the accuracies are different from reported

[anzr299]

I have tried to replicate the DCN+ model and evaluated it on the test dataset. The mAP for small objects I am getting is 17.8 unlike the 34.6 reported in the paper.

I would like to discuss if I am doing something wrong here?

[ahmedxomar101]

@anzr299 Could you please share the performance on the other metrics?

A screenshot would be optimal. Thanks.

[anzr299]

I had put it on an excel sheet since I deleted the checkpoints.

[ahmedxomar101]

@anzr299 Amazing. Thank you for sharing.

Did you change the definitions for APs, m, l in your evaluation? It is a not nice trick to increase the performance of APs, m, l on paper

Check this part in the paper

[anzr299]

Yeah, I did go through the paper and confirmed everything. I ran the same test script they gave. Also I would like to point out the part where the Learning rate in the code was 0.05 when it was mentioned in the paper to be 0.01.

[ahmedxomar101]

@anzr299 Very interesting! Would you mind if we talk privately? My LinkedIn account is listed on my profile.

[anzr299]

sure! I have sent a request.