WU Yue

@SamihaSara Heads are at https://github.com/wuyuebupt/doubleheadsrcnn/blob/master/maskrcnn_benchmark/modeling/roi_heads/box_head/box_head.py. For the second question, that's a general process in object detection. Check this function: https://github.com/wuyuebupt/doubleheadsrcnn/blob/a744b4121d52935741f49d845bae7878270ea291/maskrcnn_benchmark/modeling/roi_heads/box_head/loss.py#L52

@SamihaSara In my understanding, you will need to modify, data loader, head strucutre and loss. For data loader, https://github.com/wuyuebupt/doubleheadsrcnn/blob/master/maskrcnn_benchmark/data/datasets/coco.py#L39 For head strucutre, https://github.com/wuyuebupt/doubleheadsrcnn/blob/master/maskrcnn_benchmark/modeling/roi_heads/box_head/box_head.py For loss, https://github.com/wuyuebupt/doubleheadsrcnn/blob/master/maskrcnn_benchmark/modeling/roi_heads/box_head/inference.py https://github.com/wuyuebupt/doubleheadsrcnn/blob/master/maskrcnn_benchmark/modeling/roi_heads/box_head/loss.py Hope these help.

@yangmin666 You are right, cls_nonloca and reg_nonlocal is not used. Instead, self.shared_nonlocal is used. The code is at: https://github.com/wuyuebupt/doubleheadsrcnn/blob/master/cmd_train.sh#L35 https://github.com/wuyuebupt/doubleheadsrcnn/blob/a744b4121d52935741f49d845bae7878270ea291/maskrcnn_benchmark/modeling/roi_heads/box_head/roi_box_feature_extractors.py#L186 There are four losses in total: conv reg, conv cls,...

@userDJX I take a quick look at their results and find two potential issues that might affect the results. 1. Results from the first 20 classes are lower than our...

@userDJX Hey, I make some changes to your code and get one close result. Some implementations for the distillation are still different. But I think your implementation is kind of...

@HuangLian126 Thanks for interested in our work. PoolerNeighbor expands the proposals before ROIAlign compared to Pooler. The function is at https://github.com/wuyuebupt/doubleheadsrcnn/blob/master/maskrcnn_benchmark/modeling/poolers.py#L165, called at https://github.com/wuyuebupt/doubleheadsrcnn/blob/master/maskrcnn_benchmark/modeling/poolers.py#L201.

@SoYeol-Lee Sorry that I thought I have replied until I received the next comment. Where to use the evaluation flag can be found at the code: https://github.com/wuyuebupt/doubleheadsrcnn/blob/master/maskrcnn_benchmark/modeling/roi_heads/box_head/inference.py#L190 The mapping from...

@AliceAsky check my comment above.

@igygi Sorry about the confusion. I think you can simply clone from this repo instead, e.g. `git clone https://github.com/wuyuebupt/doubleheadsrcnn` and `cd doubleheadsrcnn`.

@brjathu The evaluation metric for MS-Celeb-1M is top-1 accuracy.