LI Wentong

@abhigoku10 For BDD100k, the methods in this [toolbox](https://github.com/LiWentomng/BoxInstSeg) can work for instance classes , like car . For the lane and road, I have not try these objects. I guess...

Hello, @zhaoyangwei123 The gt bounding box can prevent the interference from background and other objects. We achieve this goal by the binary box-based mask (`box_mask_target`) as following, ``` mask_scores_phi =...

@ccdongxu Hello, the annotaion json file needs the coco format. I upload the [coco-style json file](https://drive.google.com/file/d/1FqoHjbfduFaHUc1PHktUPLuEZp3SAp1m/view?usp=sharing) (Google driver) of VOC, you can refer to it.

@ccdongxu The training process only needs the `bbox` fields in the train.json file. The val/test process need both `segmentation` and `box` fields to evaluate the segmentatio performance (or bbox annotaion...

@TaoshuaiZ 你好，mIoU的计算可以参考此处的[代码](https://github.com/LiWentomng/Point2Mask/blob/main/datasets/coco_panoptic.py#L1991)。

@lin1061991611 你好，问题解决了吗？ 可以尝试添加 warmup的策略。 训练刚开始 让模型只采用projection loss，使其保证快速收敛 且模型收敛方向正确，训练后期 采用level set进一步的细致优化。

Hi, @lxtGH, thanks for sharing. In the section 4.4, the reference of MAL is incorrect. Please check it. Besides, our work `BoxLevelset (ECCV2022)` as a typical weakly box-supervised instance method...

@shanghangjiang Hello, I have test the default settings for boxinst. It works well. The training logs are listed as the following: 2023-03-17 20:44:14,061 - mmdet - INFO - Distributed training:...

[#9](https://github.com/LiWentomng/boxlevelset/issues/9)

@matanPercepto Hello, please try to use `pytorch1.9.0` under CUDA 11.1.1, and re-compile tree_filter_cuda operation by the following. ``` cd mmdet/ops/tree_filter python setup.py build develop ```