jshilong / grouprcnn Goto Github PK

Thank you for your share. I want to ask what is the 'sampling_results' in 'group_roi_head'?

My environment is
mmcv-full 1.5.0
mmdet 2.22.0
torch 1.12.0+cu113

occur the error in the title, can you help me?

Nice work! However， the code seems not complete.​ For example, in the paper, you use retinaHead to replace RPNHead, but, you do not offer the fixed retinaHead. Hope you can update those incomplete code. Thanks!​

The GroupRCNN is based on the Cascade RCNN.

The performance of Cascade RCNN with 100% label is about 41AP.

But the performance of GroupRCNN with 10% label is about 39.2AP, which seems too high.

For the PointDETR, the performance of teacher model with 10% label is 23.7AP, which is much lower than yours. But the gap of student performance is not as significant (teacher:39.2 vs. 23.7, student: 32.6 vs. 30.3). It seems very strange.

Thank you for sharing your great work, GroupRCNN for WSSOD.

As in the below figure, GroupRCNN is trained with a subset (10% or 20% or 50%) of COCO fully-labeled datasets.

Are there any rules or algorithms to sample a subset of COCO fully-labeled datasets?
For example, when randomly sampling 10% of fully-labeled datasets, the minor (long-tailed) classes are rarely included in the 10% sampled dataset.

Did you consider the distribution of categories when sampling the COCO dataset or do you have any special rules for sampling the dataset?

Sorry to bother you, but how to set the file parameters when using two Gpus or one GPU? Is it set directly through the command? When I tried this method, I got an error as shown in the figure below:

Hope to receive your reply, thank you!

Hello, can you give the code implementation of the semi-supervised strategy for group r-cnn in supplementary material C?

What is the training time for the dataset used?

ImportError: cannot import name 'select_single_mlvl' from 'mmdet.core

hello, I use the dataset without segmatation and olny bbox information，but the result have a wave about max-min =4MAP.Although I change the point to in the object,the result also wave.Do you have the same? PS:the result only the point to bbox generator.

Hi,
Thanks for sharing the code.
Can you clarify if the GroupRCNN approach ensures that a pseudo box is predicted for every point annotation? Can there be points for which no pseudo-box is predicted by the teacher model?

I'm sorry to bother you. I encountered a problem: Because my own data set did not have segmentation data set information, I obtained performance A after generating point annotations with box information, and then obtained performance B after manual labeling training, but performance B was less than performance A. My understanding is that manually annotated point annotations are higher than randomly generated point annotations, but the model performance has not been improved, may I ask what the possible problem is?
Looking forward to hearing from you！

Sorry to interrupt, regarding the program of the prior part, I can understand that it is to get the index of the first three points through the feature map, and get the anchor points through the index to change the template of the generated anchor box to achieve the purpose (the target of the box is consistent with the target of the point annotation). My understanding: the code group_rcnn.py> _get_dummy_bboxes_single: bboxes=self.rpn_head.bbox_coder.decode(priors, bbox_pred, max_shape=img_shape), where priors is [x, y, x,y]+base_anchors, The first part is the anchor point obtained according to the index, the second part is the base anchor box obtained according to the anchor box generator with the anchor point as (0, 0), so that a new anchor box template is established with the information of the point annotation. The proposal (without NMS and TOPK versions) is obtained by calculating the delta2bbox algorithm with bbox_pred (regression parameter obtained using rpn_head) and prior in decode.