FLM-Occ: Feed-forward Likelihood Maximization for Efficient Indoor Occupancy Prediction
FLM-Occ enhances indoor occupancy prediction efficiency via feed-forward likelihood maximization, achieving 3.7x speedup with only 32 superquadrics.
Key Findings
Methodology
This paper introduces a novel Feed-forward Likelihood Maximization (FLM) framework, redefining occupancy prediction as voxel distribution estimation. A network is trained to predict a mixture model that maximizes the likelihood over ground-truth occupied voxels in a feed-forward manner. Mixture weights are defined as normalized primitive volumes to implicitly enforce simplex constraints, and novel voxelization formulas are derived.
Key Results
- FLM-Occ achieves superior accuracy on the Occ-ScanNet dataset using only 32 superquadrics, 2.7% better than prior state-of-the-art, while running 3.7 times faster.
- Compared to EmbodiedOcc and SplatSSC, FLM-Occ achieves significantly better accuracy-efficiency trade-offs.
- FLM-Occ consistently outperforms existing methods across different numbers of primitives, especially with fewer primitives.
Significance
FLM-Occ significantly improves the efficiency and accuracy of indoor occupancy prediction, reducing the prediction of spurious primitives and enhancing both representational and computational efficiency. This method holds significant implications for academia and industry, particularly in real-time processing applications.
Technical Contribution
FLM-Occ redefines the training objective of occupancy prediction, using maximum likelihood estimation to optimize the parameters of a mixture model. This approach fundamentally differs from existing voxel classification methods, enabling distant primitive relocations and enhancing model representational capacity and computational efficiency.
Novelty
FLM-Occ is the first to reformulate occupancy prediction as voxel distribution estimation, introducing a novel feed-forward likelihood maximization framework that enables direct refinement of primitives using image features without relying on depth maps.
Limitations
- FLM-Occ may require more primitives for complex scenes, potentially affecting computational efficiency.
- The method relies on high-quality image feature extraction and may be sensitive to image quality.
Future Work
Future research directions include optimizing FLM-Occ's performance in complex scenes, exploring more types of primitives, and assessing applicability across different application scenarios.
AI Executive Summary
Recent indoor occupancy prediction methods typically use Gaussian primitives as a sparse 3D representation to enhance computational efficiency. However, these methods rely on voxel classification for training, which imposes only local constraints and lacks global supervision on the distribution of primitives. This results in the prediction of spurious primitives in empty regions, undermining both representational and computational efficiency. To address this issue, the paper introduces a novel Feed-forward Likelihood Maximization (FLM) framework, redefining occupancy prediction as voxel distribution estimation.
In FLM, a network is trained to predict a mixture model that maximizes the likelihood over ground-truth occupied voxels in a feed-forward manner. To enable end-to-end training of networks and voxelization of a standard mixture model, mixture weights are defined as normalized primitive volumes to implicitly enforce simplex constraints, and novel voxelization formulas are derived. Based on FLM, the authors propose FLM-Occ, a novel method capable of relocating randomly initialized primitives over long distances to model a scene.
On the Occ-ScanNet dataset, FLM-Occ achieves superior accuracy using only 32 superquadrics, 2.7% better than the prior state-of-the-art, while running 3.7 times faster. This method significantly improves the efficiency and accuracy of indoor occupancy prediction, reducing the prediction of spurious primitives and enhancing both representational and computational efficiency. FLM-Occ holds significant implications for academia and industry, particularly in real-time processing applications.
Deep Dive
Abstract
Recent indoor occupancy prediction methods adopt Gaussian primitives as a sparse 3D representation for computational efficiency. However, their training relies on voxel classification, which imposes only local constraints and lacks global supervision on the distribution of the primitives. Therefore, they inevitably predict spurious primitives in empty regions, undermining both representational and computational efficiency. To address this, we propose Feed-forward Likelihood Maximization (FLM), a novel framework that reformulates occupancy prediction as voxel distribution estimation. In FLM, a network is trained to predict a mixture model that maximizes the likelihood over ground-truth occupied voxels in a feed-forward manner. To enable end-to-end training of networks and voxelization of a standard mixture model, we define mixture weights as normalized primitive volumes to implicitly enforce simplex constraints and derive novel voxelization formulas. Based on FLM, our FLM-Occ, a novel method that is capable of relocating randomly initialized primitives over long distances to model a scene. On Occ-ScanNet, FLM-Occ achieves superior accuracy using only 32 superquadrics, 2.7% of the prior SoTA, while running 3.7 times faster.