Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Neural Bounding

Stephanie Wenxin Liu, Michael Fischer, Paul D. Yoo, Tobias Ritschel

TL;DR

Neural Bounding reframes geometric bounding as learning to classify space into free or occupied, enforcing zero false negatives through a dynamically weighted asymmetric loss. The method trains compact neural bounds and optionally arranges them into hierarchies with early-out capabilities, achieving significantly fewer false positives and substantial speedups in practice. It demonstrates applicability across 2D–4D queries, dynamic scenes, and high-dimensional spaces, and shows favorable trade-offs against classical bounding methods. While it does not provide formal guarantees of zero FN, it offers a practical, scalable approach to conservative bounding with real-world impact in graphics and vision pipelines.

Abstract

Bounding volumes are an established concept in computer graphics and vision tasks but have seen little change since their early inception. In this work, we study the use of neural networks as bounding volumes. Our key observation is that bounding, which so far has primarily been considered a problem of computational geometry, can be redefined as a problem of learning to classify space into free or occupied. This learning-based approach is particularly advantageous in high-dimensional spaces, such as animated scenes with complex queries, where neural networks are known to excel. However, unlocking neural bounding requires a twist: allowing -- but also limiting -- false positives, while ensuring that the number of false negatives is strictly zero. We enable such tight and conservative results using a dynamically-weighted asymmetric loss function. Our results show that our neural bounding produces up to an order of magnitude fewer false positives than traditional methods. In addition, we propose an extension of our bounding method using early exits that accelerates query speeds by 25%. We also demonstrate that our approach is applicable to non-deep learning models that train within seconds. Our project page is at: https://wenxin-liu.github.io/neural_bounding/.

Neural Bounding

TL;DR

Neural Bounding reframes geometric bounding as learning to classify space into free or occupied, enforcing zero false negatives through a dynamically weighted asymmetric loss. The method trains compact neural bounds and optionally arranges them into hierarchies with early-out capabilities, achieving significantly fewer false positives and substantial speedups in practice. It demonstrates applicability across 2D–4D queries, dynamic scenes, and high-dimensional spaces, and shows favorable trade-offs against classical bounding methods. While it does not provide formal guarantees of zero FN, it offers a practical, scalable approach to conservative bounding with real-world impact in graphics and vision pipelines.

Abstract

Bounding volumes are an established concept in computer graphics and vision tasks but have seen little change since their early inception. In this work, we study the use of neural networks as bounding volumes. Our key observation is that bounding, which so far has primarily been considered a problem of computational geometry, can be redefined as a problem of learning to classify space into free or occupied. This learning-based approach is particularly advantageous in high-dimensional spaces, such as animated scenes with complex queries, where neural networks are known to excel. However, unlocking neural bounding requires a twist: allowing -- but also limiting -- false positives, while ensuring that the number of false negatives is strictly zero. We enable such tight and conservative results using a dynamically-weighted asymmetric loss function. Our results show that our neural bounding produces up to an order of magnitude fewer false positives than traditional methods. In addition, we propose an extension of our bounding method using early exits that accelerates query speeds by 25%. We also demonstrate that our approach is applicable to non-deep learning models that train within seconds. Our project page is at: https://wenxin-liu.github.io/neural_bounding/.
Paper Structure (25 sections, 6 equations, 12 figures, 2 tables, 1 algorithm)

This paper contains 25 sections, 6 equations, 12 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Previous work
  3. Our Approach
  4. Method
  5. Neural Bounding Hierarchies
  6. Neural Early-out
  7. Implementation
  8. Evaluation
  9. Methods
  10. Tasks
  11. Indicators
  12. Query types
  13. Metrics
  14. Tightness
  15. Speed
  16. ...and 10 more sections

Figures (12)

  • Figure 1: Motivating examples for conservative queries: we encode the world as a NN trained with (left) and without (right) our proposed approach. When used for 2D path-planning, our representation achieves a collision-free trajectory (left), whereas the non-conservative NN collides (arrows). Similar observations can be made when using the NN as a proxy for 3D cloth collision leading to erroneous interpenetration in the middle row, or as a proxy to cull shadow rays where the stings are missing in the shadow.
  • Figure 2: Result for OurkDOP. Image credit: 588ku, PNGTree.com
  • Figure 3: Conservative boun-ding of never- (blue) and always-hit (yellow).
  • Figure 4: A hierarchy of neural bounding networks. The first level is show in blue, the two child nodes in yellow and the leafs in pink. Note that it is not required for the higher-level bounding to bound the lower-level boundings. It is only required to bound the indicator.
  • Figure 5: FP rates (vertical axis) for each test object, sorted in ascending order (horizontal axis). We show all methods (colored lines) across all dimensions (horizontal subplots). While Tbl. \ref{['tab:Results']} reports average performance, this graphical representation reveals that our method OurNN, despite having a few higher negative examples, consistently yields the lowest FP rates across all categories. The average of OurkDOP and OurNNEarly seem affected by some hard cases.
  • ...and 7 more figures