train blender datas using ./script/benchmarking/launch_train_blender.sh and export pointcloud, but the pointcloud output is incorrect and has many white points.

[annci62]

Describe the bug
A clear and concise description of what the bug is.

To Reproduce
Steps to reproduce the behavior:
ns-train Lego in nerf_synthetic and ns-rander, but output is error, but training Lego using instant-gnp and output is right

To Reproduce
Steps to reproduce the behavior:
1)ns-train instant-ngp --data=./data/blender/drums --experiment-name=blender_drums_2023-04-25 --relative-model-dir=nerfstudio_models/ --steps-per-save=1000 --max-num-iterations=16500 --logging.local-writer.enable=False --logging.enable-profiler=False --vis tensorboard blender-data & GPU_PID[1]=316294 --pipeline.model.background-color white --pipeline.model.proposal-initial-sampler uniform --pipeline.model.near-plane 2. --pipeline.model.far-plane 6. --pipeline.datamanager.camera-optimizer.mode off --pipeline.model.use-average-appearance-embedding False --pipeline.model.distortion-loss-mult 0 --pipeline.model.disable-scene-contraction True
2)ns-export pointcloud --load-config ./outputs/blender_drums_2023-04-25/instant-ngp/2023-04-25_164500/config.yml --output-dir ./outputs/blender_drums_2023-04-25/instant-ngp/2023-04-25_164500/ --num_rays_per_batch 16384

The pointcloud output is

[MvWouden]

I'm assuming this has to do with the flag you added --pipeline.model.background-color white. If you have png's with transparent backgrounds, the background colour is generally randomized so that no background colour is learned. By fixing it to be white, the NeRF model learns a white background, hence why the white points show up in your example.

You could retry without this flag to see if it makes a difference for you.

[annci62]

According to #806 (comment), added "--pipeline.model.background-color white". If removing "--pipeline.model.background-color white", train is ok, but the progress of exporting pointcloud always is 0% and can NOT end.

[ATing0203]

hi! have you solved this problem yet? @annci62

[annci62]

@ATing0203 add bounding box