clipgstream-dynamic-scene-reconstruction

Train the reference clip first with trainReferenceClip.py, then run trainSourceClip.py for each subsequent clip using the same output directory. Each source clip can run independently on a different GPU via CUDA_VISIBLE_DEVICES for parallel training.

Run render.py with the same project_total_frames, clip_size, and config used during training, plus --skip_train to render only test views. Outputs are written to the test/renders directory, and images2video.py compiles them into videos.

ClipGStream expects per-frame directories containing COLMAP-format sparse camera parameters and undistorted images, plus a plys folder with the reference clip point cloud and residual point clouds per source clip. A preprocessing pipeline in data_process/custom_dataset generates this from raw videos.

Flickering happens when source clips do not inherit static information from the reference clip or use mismatched output paths or configs. Ensure all clips share the same -m model directory and config file, and consider 2-3 frame overlap between clips.

Reduce the clip_size to split the sequence into smaller clips, lower log2_hashmap_size from 19 to 18 in the config, or reduce batch size. These changes decrease GPU memory consumption at the cost of more clips or slightly reduced capacity.