Guangxing Han

Yes, this is same to me. The official released model have better accuracy about 55% mAP on novel class. But if I train it by myself, the result is only...

Thank you @PeizeSun for your quick response and the wonderful codes. If I understand correctly, you concat the RoI features in the last stage to the current stage for classification....

> Hi Bruno .. i would say the objectives are 100% the same. So its better to go with a Microsoft research paper that has code rather than some random...

> Hi Mr Han .. no, i havent reproduced the results because i would like to train this on charts / graph data exclusively and im preparing the datasets ..having...

Sorry for the late reply. The error is because self.support_df does not find the bounding box with id defined by dataset_dict['annotations'][0]['id']. But I do not know why it happens. You...

Feel free to contact me by email ([email protected]) or open issues here for any technical issues. I will check them regularly. Thanks.

The reason might be that the program does not find the original jpg files. Can you double confirm that the two above image files are good?

1. Yes, coco/support/trainval2014 is created by the script. 2. As far as I know, lots of people have used this projects so far, and I did not know any problem...

We followed the previous work (TFA, ICML 2020) to prepare the dataset (https://github.com/ucbdrive/few-shot-object-detection/tree/master/datasets). Please see below for an overview of the coco folder. ``` coco/ annotations/ trainvalno5k.json instances_val2014.json # same...

No, they should not be deleted. We do not save the original images anywhere else with our scripts.