Feature Splatting

This repo implements feature splattiing, which combines gaussian splatting with feature distillation. Compared to simple extension to the original gaussian splatting, our implementation is much faster and more memory-efficient.

For plain 768-dim feature rendering, our method achieves ~60% speedup with negligible accuracy loss (FP32->FP16). For further speedup, we render implicit 32-dim features and then convolve them to 768-dim features. This achieves additional ~10x speedup.

Setup dependency

Clone the codebase.

cd ~
git clone --recursive https://github.com/vuer-ai/feature-splatting-inria

Follow INSTALL.md to install required dependencies.

Prepare data

Currently, the repo supports training from the following data formats

Colmap-processed RGB images (defined by a colmap database and a folder of images)
Synthetic data (defined by a transforms.json and images)

For testing the codebase, we recommend donwloading sample datasets (kitchen bulldozer, garden vase) and following the instructions below.

# Assume that garden_table_colmap.zip was downloaded to ~/Downloads and feature-splatting-inria is under home directory
cd ~
cd feature-splatting-inria
mkdir feat_data && cd feat_data
mv ~/Downloads/garden_table_colmap.zip ./
unzip garden_table_colmap.zip
cd ..

Alternatively, you can follow the instructions here to use colmap to process your own data or generate synthetic datasets from objaverse.

Computing features

For efficient feature computation, the released code makes a trade-off with the original paper - instead of using SAM, we use MobileSAMV2 to generate object-level masks. This drastically improves the feature extraction speed.

We have automated the weight download of MaskCLIP and MobileSAMV2 for you. To extract features, just run

python compute_obj_part_feature.py -s feat_data/garden_table

Training

To train feature splatting on the sample dataset, you can run

python train.py -s feat_data/garden_table -m output/garden_table --iterations 10000

Rendering

After the training is done, you can render images via,

python render.py -m output/garden_table -s feat_data/garden_table --camera_slerp_list 0 1 --with_feat --clip_feat --text_query 'a vase with flowers' --step_size 10

In this example, the results will be available under output/garden_table/interpolating_camera/ours_10000/renders. The added options render latent features and CLIP heatmaps in response to the specified word. To visualize latent features via PCA, you can run

python pca_feature_viz.py --input_dir output/garden_table/interpolating_camera/ours_10000/renders --output_dir output/garden_table/interpolating_camera/ours_10000/pca_renders

and the PCA visualized images will be available in the specified output directory.

Physics Simulation

The code for physical simulation is done in two steps. The first step is to get a 3D segmentation of the target object (with optional rigid part), and the second step is to actually run the MPM physics engine.

To reproduce the vase+flower example, use the following segmentation commands.

python segment.py -m output/garden_table --final_noise_filtering --inward_bbox_offset 0.15 --rigid_object_name vase --interactive_viz

Though feature splatting has a clear object boundary due to the use of SAM, in reality, the per-gaussian features are often noisy. Simple CLIP GS-text feature cosine similarity thresholding works for most of the Gaussians, but there are often some noises. Hence, we implement several post-processing techniques in segment.py.

Empirically, for selecting objects using your own data, here are the important ones

Text prompting

--fg_obj_list: text prompts for positive objects, separated by commas.
--bg_obj_list: text prompts for background objects (object near FG object that needs to be excluded), separated by commas
--ground_plane_name: a single text string for ground plane contact with the object (e.g., tabletop, floor).
--rigid_object_name (Optional): given a foreground object, select high-confidence rigid particles inside

Post processing

--object_select_eps (float): EPS for the DBSCAN clustering algorithm to select a single primary object. If some parts of the object are missing -> increase EPS. If it includes unwanted parts of other similar objects -> decrease EPS. This may require some tweaking due to the scale ambuiguity of colmap.
--inward_bbox_offset (float, Optional): This function selects GS within a ground-aligned bbox. Larger offset = more conservative bbox. Smaller offset = more aggresive bbox.
--final_noise_filtering (bool, Optional): If True, do an inverse KNN dilation to filter out final noises around the object.

In addition, we alos offer an option --interactive_viz to visualize the effects of every post-processing methods.

After the object is segmented, run

python mpm_physics.py -m output/garden_table --rigid_speed 0.3 --use_rigidity

which would compute the per-particle trajectory of selected Gaussians. To render the physical movement, run,

python render.py -m output/garden_table -s feat_data/garden_table --camera_slerp_list 54 58 --step_size 500 --with_editing

The output is also located at output/garden_table/interpolating_camera/ours_10000/renders.

TODOs

Add instructions for serializing synthetic assets from objaverse
Support more interactive GS selection process with Vuer / supersplat canvas
Experiment with projection-based method that would remove need for post-processing
Organize code for rotation estimation
Multi-scale CLIP features
Add support for sparse graident computation in custom pytorch Ops to save GPU memory.