search

gaobb / DCFS

[NeurIPS 2022] Decoupling Classifier for Boosting Few-shot Object Detection and Instance Segmentation
https://csgaobb.github.io/Projects/DCFS
MIT License
29 stars 2 forks source link

configs/coco/dcfs_FSIS_r101_novel_shot_seedx.yaml. #3

Closed jinse-zyy closed 1 year ago

jinse-zyy commented 1 year ago

Hello, I couldn't find "configs/coco/dcfs_FSIS_r101_novel_1shot_seedx.yaml." Could you please provide it to me? Thank you.

gaobb commented 1 year ago

hi，

We thank you for your attention to our work.

This issue has been fixed by updating config files for training FSIS and gFSIS in the "run_coco_fsis.sh". In fact, the same config file is used to train FSOD and FSIS (gFSOD and gFSIS) in our implements.

If there is anything, please do not hesitate to let me know.

jinse-zyy commented 1 year ago

You're very welcome! While running the experiments, I also noticed some configuration issues in the run_coco_fsod.sh script: 1. The --eval-only option on line 48 is redundant. 2. You need to add the following line of configuration separately for fsod and gfsod: MODEL.ROI_BOX_HEAD.BBOX_CLS_LOSS_TYPE ${classloss}.

In addition, regarding the experimental section in the paper, I have a question about the distinction between "Ours" and "Ours." You mentioned that "Ours" is the average of 10 runs, so what does "Ours" represent? Is it the value for the first run?

gaobb commented 1 year ago

  1. We train and test simultaneously during the training process, so the test file is not redundant. https://github.com/gaobb/DCFS/blob/07cad8d0e8ed4239cebb0d1428963264de80241b/run_coco_fsod.sh#L48

  2. The loss function is valid using the current command line mode. For example, https://github.com/gaobb/DCFS/blob/07cad8d0e8ed4239cebb0d1428963264de80241b/run_coco_fsod.sh#L46

  3. In the Experimental Details on Page 7 (or the caption of Table 3 on page 8)of our paper, there are detailed descriptions of Ours and Ours*.

Following the previous work [33], all experimental results are averaged over 10 seeds. For a fair comparison with DeFRCN [28], we also report the average results over 10 times repeated runs on seed0.

The superscript * indicates that the results are averaged over 10 times repeated runs on seed0, the same as below.

Here, Ours means that the results are averaged over 10 seeds, while Ours* implies the result is 10 times repeated runs only on seed0.

jinse-zyy commented 1 year ago

Thank you for your response. My previous questions contained errors. The incorrect configuration file is run_voc_fsod.sh.

gaobb commented 1 year ago

Thank you for your response. My previous questions contained errors. The incorrect configuration file is run_voc_fsod.sh.

We greatly appreciate you pointing out this issue, and it has been fixed in run_voc_fsod.sh.