Open Fan-QY opened 2 years ago
thangvubk
commented
2 years ago Hi. in my experiments, i also find that training without class weight achieves at least comparable semantic segmentation performance to with class weight. I think your results are normal.
Fan-QY
commented
2 years ago Hi. in my experiments, i also find that training without class weight achieves at least comparable semantic segmentation performance to with class weight. I think your results are normal.
Thank you for your reply. But when I use the backbone results without weights for further training, the results are very poor and an error occurs at the 9th epoch. Do you have any idea what could be the problem? And am I setting the weights the right way?
Fan-QY
commented
2 years ago Hi. in my experiments, i also find that training without class weight achieves at least comparable semantic segmentation performance to with class weight. I think your results are normal.
And how should I improve the training results of the backbone network now? Thanks so much!
thangvubk
commented
2 years ago Did you pretrain the backbone before training the whole network? And what are semantic mIoU and offset MAE of pretrained model?
Fan-QY
commented
2 years ago Did you pretrain the backbone before training the whole network? And what are semantic mIoU and offset MAE of pretrained model? Yes, I have completed the backbone training, which semantic mIoU Ist 38.0， offset MAE ist 3.000
2022-08-24 08:59:41,476 - INFO - Evaluate semantic segmentation and offset MAE 2022-08-24 08:59:45,174 - INFO - Class-wise mIoU: 40.5 74.9 21.7 0.1 97.4 22.8 0.0 38.9 2022-08-24 08:59:45,174 - INFO - mIoU: 37.0 2022-08-24 08:59:46,006 - INFO - Acc: 96.6 2022-08-24 08:59:52,687 - INFO - Offset MAE: 3.000
Hello, I am to want to use Softroup for instance segmentation in my own dataset. Among other things I have transformed my dataset according to S3DIS. My PC ist TitanXP (12GB) graphics and my original dataset has only 5 scenes (5 files for training). Since my individual scenes are too large and complex, I sliced them, where each slice has 4 million points. This resulted in 44 training files. The dataset has a large gap between the points in different classes and is not evenly distributed. The number of training points and instances in different categories for the 44 training files are as follows:
Total number of points of each class: [10152406, 39686946, 9433358, 1552109, 92844212, 2696146, 498094, 7429789] Under each class is the number of instances: [139, 197, 433, 36, 483, 16, 92, 41] class_numpoint_mean [73038, 201456, 21786, 43114, 192224, 168509, 5414, 181214]
My backbone's training configuration file is as follows
(The file of pretrain is the result of my 12 classes under the same dataset, and the IoU of 4 classes is 0.0% due to the uneven rhyme division of its 12 class points. So I changed the dataset from 12 class datasets to 8 class datasets ) The best results obtained are shown below, two of the categories were not identified at all.
By understanding the weights, I determined the weights by referring to the method of the website below. https://www.analyticsvidhya.com/blog/2020/10/improve-class-imbalance-class-weights/ wj=n_samples / (n_classes * n_samplesj)
The new training configuration was set up from scratch, as follows.
The best result is shown below, and the result is worse than the first backbone training.
I used the best result of the first training (becaus more mIoU) for the instance training, and I think the model cannot learn anything and generates the following error at the end of the training.
How can I improve this result and solve the problem? Thank you very much!