Questions about COCO experiments

[XinyuJiang]

1. I notice that in fine-tune stage, the COCO dataset only have one instance per image, despite of there may be number of potential instances, like some image contained chairs, and these instances without annotations would be treated as False Positive in this stage. Does it make sense?
2. In configs/COCO/R-101/10shot_baseline.yml, there have no config about "ContrastiveROIHeads", which was set in PACSCAL_VOC/split*/*shot_CL_IoU.yml. But I use this coco baseline achieve 11.6 nAP in 10-shot coco setting, and when I add config about CL in COCO config file it only achieve 10.16nAP. Whether I misunderstand something important? Waiting for your response. Thanks.

[Wei-i]

I have the same issue with (2). It seems that contrastive loss could harm coco-few-shot but benefit voc-few-shot experiments.

[Ze-Yang]

1. I notice that in fine-tune stage, the COCO dataset only have one instance per image, despite of there may be number of potential instances, like some image contained chairs, and these instances without annotations would be treated as False Positive in this stage. Does it make sense?
2. In configs/COCO/R-101/10shot_baseline.yml, there have no config about "ContrastiveROIHeads", which was set in PACSCAL_VOC/split*/*shot_CL_IoU.yml. But I use this coco baseline achieve 11.6 nAP in 10-shot coco setting, and when I add config about CL in COCO config file it only achieve 10.16nAP. Whether I misunderstand something important? Waiting for your response. Thanks.

Hi, could you share your results on COCO 10shot and 30shot after adding CL with the format below for both base and novel classes? Thanks.

AP	AP50	AP75	APs	APm	APl	AR	AR50	AR75	ARs	ARm	ARl