sankin

The boxinst results on cityscapes. 

@tianzhi0549 The CondInst result on Cityscapes(without coco pretrained) 

@tianzhi0549 Thankyou, I'll trying the coco-pretraining. I am also confused about the results without pre-training, such results are normal or not？I have found a supplenmenray material for CondInst on Cityscapes，my...

@tianzhi0549 Thanks for your responses! I also follow the settings in https://github.com/facebookresearch/detectron2/blob/master/configs/Cityscapes/mask_rcnn_R_50_FPN.yaml in my experiments. I will do some experiments with the coco pre-training, then share with you, i look...

Here is my conda environment configuration: 1. [conda_list](https://drive.google.com/file/d/1So4EqnsINUKDnWSTPyKbQUxVenHFYNMT/view?usp=drive_link) 2. [pip_list](https://drive.google.com/file/d/1mdk0JHmTpum-VuuL6wILhgdqhmEuGOIm/view?usp=drive_link) I hope it will solve your problems.

Please refer to [line557](https://github.com/PJLab-ADG/LoGoNet/blob/main/detection/al3d_det/models/modules/roi_heads/LoGoHead_kitti.py#L577) in LoGoHead_kitti.py and [line134](https://github.com/PJLab-ADG/LoGoNet/blob/main/detection/al3d_det/utils/attention_utils.py#L134) in attention_utils.py.

Have you been able to successfully replicate the two-stage performance? I’ve discovered that utilizing the ROI head for refinement results in a significant decrease in performance on the Nuscenes dataset.

How about the performance when not using MVP.