Added KDTree support for KNN. Added tests/time benchmarks for new knn model.#26
Open
finnBsch wants to merge 1 commit intoSpatialVision:mainfrom
Open
Added KDTree support for KNN. Added tests/time benchmarks for new knn model.#26finnBsch wants to merge 1 commit intoSpatialVision:mainfrom
finnBsch wants to merge 1 commit intoSpatialVision:mainfrom
Conversation
|
Thanks for the work! I used it and was able to obtain comparable results for hundreds of images in my reconstruction case. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Add KD-Tree Alternative for KNN Alignment
Overview
This PR introduces an alternative implementation for K-Nearest Neighbors (KNN) alignment in the depth completion pipeline using KD-trees, using torch_kdtree This is to improve inference time. Users can now choose between the existing
torch_clusterimplementation and the new KD-tree approach.Changes
Core Implementation
_knn_aligns_kdtree()method as an alternative to_knn_aligns_torch_cluster()knn_aligns()to dispatch between implementations based on thekd_treeparameterforward()andkss_completer()methods to accept thekd_treeboolean parameterThis change is fully backward compatible. Existing code will continue to work unchanged, if setting the
kd_treeparameter to false.API Changes
Performance Characteristics
Performance varies based on image resolution and sparsity levels. The KD-tree implementation can provide substantial speedups for high-resolution images:
Example Performance (1920x1280 resolution):
However, for smaller images,
torch_clustermay be faster.Validation and Testing
Benchmark Script
Benchmark script
benchmark_knn_direct.pyis included to measure execution time and correctness validation.Usage:
python benchmark_knn_direct.py \ --sparsity 0.60 0.90 \ --runs 5 \ --K 5 \ --height 1280 \ --width 1920Parameters:
--sparsity: List of sparsity ratios to test (default: [0.60, 0.90])--runs: Number of benchmark runs per test (default: 5)--K: Number of nearest neighbors (default: 5)--height: Image height for testing (default: 1280)--width: Image width for testing (default: 1920)Correctness Validation
The benchmark includes distance validation to ensure both implementations produce mathematically equivalent results. Any differences are due to legitimate tie-breaking when multiple points are equidistant, which can lead to different neighbors being returned. Hence, the correctness tests check for three conditions:
Dependencies
The KD-tree implementation requires:
Or install from source, which is what I did and tested with, and which is why I have not added it to the requirements.
Use the provided benchmark script to make an informed decision for your use case.
Addresses #4, #22 , and maybe #23.
cc @Ammoniro @HeoeH