In COCO, I Get a very low AP score, how can I fix it?

[sixone-Jiang]

I get less than 3 AP in COCO, but I get the AP that similar to you in VOC; I check my output log find that many object in COCO get 0 AP; how can I fix it?

[sixone-Jiang]

[03/13 11:19:10 fewx.evaluation.coco_evaluation]: Evaluation results for Non VOC 60 categories =======> AP : 0.00 [03/13 11:19:10 fewx.evaluation.coco_evaluation]: Evaluation results for Non VOC 60 categories =======> AP50: 0.00 [03/13 11:19:10 fewx.evaluation.coco_evaluation]: Evaluation results for Non VOC 60 categories =======> AP75: 0.00 [03/13 11:19:10 fewx.evaluation.coco_evaluation]: Evaluation results for Non VOC 60 categories =======> APs : 0.00 [03/13 11:19:10 fewx.evaluation.coco_evaluation]: Evaluation results for Non VOC 60 categories =======> APm : 0.00 [03/13 11:19:10 fewx.evaluation.coco_evaluation]: Evaluation results for Non VOC 60 categories =======> APl : 0.00

[Jaraxxus-Me]

How about the novel classes? The test only counts novel classes

[sixone-Jiang]

In COCO, all novel classes get 0 AP, but in VOC, I can get similar AP like you. May be I split the datasets wrong?

[Jaraxxus-Me]

Could be, you can check the steps in data generation phase

[Jaraxxus-Me]

Also check the ID mapper is correct for the VOC classes in COCO

[sixone-Jiang]

Thanks for your reply! At your reminder, I thought that non_voc60 should get 0AP in the testing stage; After the pre-training model you provided is fine-tuned, the category of nonVoc will not be recognized.

[sixone-Jiang]

[Q12seven]

May I ask if this result is obtained after fine-tuning? I can't fine-tune the code once it's running