Research map / 3D vision / Depth and stereo
Query Quantized Neural SLAM
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes query quantized neural SLAM, which quantizes coordinates, positional encodings, geometry features and TSDF priors into discrete codes so neural SDF/color networks overfit each frame faster, achieving SOTA reconstruction and tracking on RGBD benchmarks.
Problem
Neural implicit SLAM methods underfit within the few optimization iterations allowed per frame because continuous queries have large input variation, causing camera tracking drift and reconstruction artifacts.
Approach
Quantize each continuous query (coordinate, positional encoding, hash-grid geometry feature, TSDF prior) into a discrete code from a learnable codebook via nearest-neighbor VQ, so the network sees a finite set of inputs and overfits faster across frames; stabilize with Bernoulli codebook initialization, stop-gradient codebook losses, empty-space loss, and tanh augmentation of the TSDF prior.
Key results
- On Replica, averaged completion drops to 1.93 cm and completion ratio rises to 94.38%, versus 2.08 cm and 93.44% for Co-SLAM.
- On ScanNet tracking (4 scenes), average ATE RMSE improves to 8.49 cm versus 9.37 cm for Co-SLAM.
- On TUM RGB-D fr1/desk and fr2/xyz, ATE RMSE is 2.61 cm and 1.7 cm, slightly better than Co-SLAM (2.7 and 1.9); on fr3/office Co-SLAM is better (2.67 vs 2.70).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| quantize neural queries | Query Quantized Neural SLAM · Replica | Completion (cm): 2.08 -> 1.93 | Helped |
| quantize neural queries | Query Quantized Neural SLAM · Replica | ATE RMSE (cm): 1.10 -> 0.91 | Helped |
| quantize neural queries | Query Quantized Neural SLAM · ScanNet | ATE RMSE (cm): 9.37 -> 8.49 | Helped |
| replace coordinate quantization with continuous coordinates (ablation) | Query Quantized Neural SLAM · ScanNet | ATE RMSE (cm): 8.54 -> 9.15 | Hurt |
| replace the geometry-feature codebook with continuous geometry features (ablation) | Query Quantized Neural SLAM · ScanNet | ATE RMSE (cm): 8.54 -> 10.49 | Hurt |
| remove the TSDF geometry prior (ablation) | Query Quantized Neural SLAM · ScanNet | ATE RMSE (cm): 8.54 -> 9.87 | Hurt |
| replace the quantized TSDF prior with a continuous geometry prior (ablation) | Query Quantized Neural SLAM · ScanNet | ATE RMSE (cm): 8.54 -> 9.09 | Hurt |
| remove tanh augmentation of the TSDF interpolation (ablation) | Query Quantized Neural SLAM · ScanNet | ATE RMSE (cm): 8.54 -> 10.48 | Hurt |
| replace Bernoulli codebook initialization with uniform initialization (ablation) | Query Quantized Neural SLAM · ScanNet | ATE RMSE (cm): 8.54 -> 9.80 | Hurt |
| use 64 codes instead of 128 (ablation) | Query Quantized Neural SLAM · ScanNet | ATE RMSE (cm): 8.54 -> 10.53 | Hurt |
| use 256 codes instead of 128 (ablation) | Query Quantized Neural SLAM · ScanNet | ATE RMSE (cm): 8.54 -> 10.26 | Hurt |
| add depth rendering loss to RGB rendering loss (ablation) | Query Quantized Neural SLAM · ScanNet | ATE RMSE (cm): 148.9 -> 133.39 | Helped |
Surprising findings
- Naively feeding the TSDF geometry prior as input degrades performance because the network learns a shortcut mapping prior to output; a tanh transform fixes this.
- Uniform or Gaussian codebook initialization causes unstable optimization, while Bernoulli initialization is needed for stability.
Limitations
- Reconstruction on some scenes (e.g., ScanNet Acc.) is worse than baselines such as NICE-SLAM and Co-SLAM
- Performance is sensitive to the number of codes (64 and 256 both degrade vs 128)
- Tracking on some TUM and ScanNet scenes is slightly worse than Co-SLAM or NICE-SLAM
Experimental setup
Datasets: Replica ScanNet SyntheticRGBD TUM RGB-D NICE-SLAM ScanNet subset with BundleFusion poses
Benchmarks: Replica ScanNet SyntheticRGBD TUM RGB-D
Metrics: Depth L1 Accuracy Completion Completion Ratio ATE RMSE
Baselines: iMAP NICE-SLAM NICER-SLAM Co-SLAM DF Prior Go-Surf
Builds on: Co-SLAM NICE-SLAM VQ-VAE InstantNGP DF Prior NeuS
Closest papers by meaning
- SNI-SLAM: Semantic Neural Implicit SLAM (CVPR 2024 · similarity 0.87)
- Benchmarking Implicit Neural Representation and Geometric Rendering in Real-Time RGB-D SLAM (CVPR 2024 · similarity 0.86)
- Loopy-SLAM: Dense Neural SLAM with Loop Closures (CVPR 2024 · similarity 0.86)
- DeepPointMap: Advancing LiDAR SLAM with Unified Neural Descriptors (AAAI 2024 · similarity 0.85)
- MNE-SLAM: Multi-Agent Neural SLAM for Mobile Robots (CVPR 2025 · similarity 0.85)
- MAGiC-SLAM: Multi-Agent Gaussian Globally Consistent SLAM (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.