Xiaofeng Wang (Jeff)

Hi, I came across the same problem, but it seems that it is not a bug. Because the "im2col_step" is a parameter. As is shown bellow: We can always choose...

To boost the performance, MVSTER* is trained on 1600 x1200 and then tested on 1600 x1200. In the ablation, we train on 640 x 512 and test on 864 x...

It is strange that using bs=4@2GPU produces worse results than bs=2@4GPU. Here are some suggestions to deal with the problem: (1) Use sync_bn (torch.nn.SyncBatchNorm.convert_sync_batchnorm). (2) Carefully tune the LR, our...

> 1. "mvsnet_xxx_l3.ply" is the **final** point cloud output of the scan xxx, and other ply files (e.g., scan1/ply_local/00000000.ply) are the intermediate point clouds for simple visualization. To be more...

You should change the BN configurations at [here](https://github.com/JeffWang987/OpenOccupancy/blob/a97019a31b968e25206472eb28d10ee0ec0ab9b5/projects/configs/baselines/CAM-R50_img1600_128x128x10.py#L80) (SyncBN-->BN)

When visible_mask=Ture, we calculate the visible region (of camera/LiDAR), which is only used for visualization. Note that the visibility code is still under construction, and we do not use it...

We might leverage the mask later. In our early-stage experiment, we found that the mIoU of visible regions is much higher than that of invisible areas. And we encourage the...

> errors in @JeffWang987 > > ``` > bash tools/dist_test.sh ./projects/configs/baselines/CAM-R50_img1600_128x128x10.py \ > ./work_dirs/CAM-R50_img1600_128x128x10/best_SSC_mean_epoch_23.pth 8 \ > --show --show-dir ./work_dirs/CAM-R50_img1600_128x128x10/show > ``` > > ``` > File "/mnt/intel/jupyterhub/qihao/sync_projects/3_occupancy/OpenOccupancy/projects/occ_plugin/occupancy/apis/test.py", line 142,...

> Also, how to visualize GT voxel in `OpenOccupancy/data/nuScenes-Occupancy/v0.0-base/scene_aa45d94630be481787f20121ce89a605/occupancy/548ca66e9adc48ca88d1bcaeeafc35da.npy` Thanks! > > shape: (136774, 4) The Visualization code is being reframed (might be released in the future), you can DIY...

> 在 4090 8卡机上跑的 (环境就用作者提供的 pytorch 版本), 不过这 loss 看起来很异常: > > ``` > ^@2024-07-17 17:50:37,740 - mmdet - INFO - Epoch [1][600/3517] lr: 3.000e-04, eta: 14:55:15, time: 1.008, data_time:...