Yikai Wang

Hi, thanks for your interest. segmentation dataset: https://drive.google.com/drive/folders/1mXmOXVsd5l9-gYHk92Wpn6AcKAbE0m3X image translation dataset: https://github.com/alexsax/taskonomy-sample-model-1 Both segmentation and image translation codes provide train and val splits:(https://github.com/yikaiw/CEN/tree/master/semantic_segmentation/data/nyudv2, https://github.com/yikaiw/CEN/tree/master/image2image_translation/data).

It is not necessary. You can place the dataset folder (that contains "depth", "masks" and "rgb") to any path, as long as you modify the data path in https://github.com/yikaiw/CEN/blob/40f277ed1a377a3c81f979a6c534ae268773aa9d/semantic_segmentation/config.py#L5

代码中分别对rgb，depth和ens的loss求和，主要为了解耦ens weight和网络的学习过程，即在更新ens weight时网络参数不更新。 https://github.com/yikaiw/CEN/blob/40f277ed1a377a3c81f979a6c534ae268773aa9d/semantic_segmentation/models/model.py#L316

Hi, do you mean SUN RGBD for semantic segmentation or SUN RGBD for 3D object detection?

1e-6 empirically works better. In fact, choosing this hyper-parameter is not strict, as long as we keep the final exchanged ratios around 30%~50%.

Download address: https://pan.baidu.com/s/1LDZA6d-3eQgpZpR5hONN2w Extract code: pike

The file 'all_obbs2d_modified_nearest_has_empty.pkl' is generated with the same method as 'all_obbs_modified_nearest_has_empty.pkl', but stores 2D bboxes instead of 3D bboxes. 'all_obbs_modified_nearest_has_empty.pkl' is generated in https://github.com/yikaiw/TokenFusion/blob/3834ccf7765bb0bd50ea729069ad5adbd6de288d/object-detection-3d/sunrgbd/sunrgbd_data.py#L342-L345

same issue. where to find data.himv?

Thanks for your recognition. We use 8 V100 GPUs for training ResNet50 and ResNet18. Actually, 4 GTX1080 GPUs are enough for ResNet18.

On 8 V100 GPUs, ResNet50 needs 4 days, and ResNet18 only needs 1 day. On 4 GTX1080 GPUs, ResNet18 needs about 2 days. Note that ImageNet data should be stored...