sharinka0715
commented
4 months ago Hi,
The 3D distillation experiments should be conducted on a scene dataset rather than a single scene. If you just want to try our semantic gaussians on a single scene, you do not need to conduct 3D distillation. You can just run view_viser.py to view performances on 2D projection results.
I run the distill.py after have the fusion result, in /data/new_disk4/shenzhh/semantic-gaussians/fusion/0.pt, but the code went wrong.
Traceback (most recent call last): File "/data/new_disk4/shenzhh/semantic-gaussians/distill.py", line 243, in
distill(config)
File "/data/new_disk4/shenzhh/semantic-gaussians/distill.py", line 75, in distill
dataset = FeatureDataset(
File "/data/new_disk4/shenzhh/semantic-gaussians/dataset/feature_dataset.py", line 34, in init
features = os.listdir(os.path.join(point_dir, scene))
FileNotFoundError: [Errno 2] No such file or directory: '/data/new_disk4/shenzhh/semantic-gaussians/fusion/cameras.json'
Since the edited config is given below: scene: scene_path: "/data/new_disk4/scene20240201/data1/1150/image_undistortion_white" dataset_name: "cocomap" test_cameras: False colmap_images: "images" colmap_eval_hold: 8 downscale_ratio: 1 white_background: True device: "cuda:3"
pipeline: convert_shs_python: False compute_cov3d_python: False debug: False seed: 1
model: sh_degree: 3 model_dir: "/data/new_disk4/scene20240201/data1/1150/output" dynamic: False load_iteration: 30000 device: "cuda:3"
fusion: out_dir: "/data/new_disk4/shenzhh/semantic-gaussians/fusion"
distill: exp_name: openseg_new model_3d: MinkUNet34A voxel_size: 0.02 aug: True feature_type: all lr: 0.001 epochs: 100 loss_type: cosine schedule_milestones: [20, 40, 60, 80, 100] schedule_gamma: 0.3 batch_size: 1 num_workers: 16 test_interval: 10 save_interval: 10