As of June 2025, are there any visual odometry systems using end-to-end deep learning…
- User
As of June 2025, are there any visual odometry systems using end-to-end deep learning that outperform extracting correspondences with deep learning and using classical visual odometry chains (i.e. epipolar error minimization)?
Assistant
Based on my knowledge through January 2025, the state of visual odometry (VO) was still evolving, with hybrid approaches generally outperforming pure end-to-end methods for most practical applications.
The classical pipeline approach - using deep learning for feature extraction/matching followed by geometric optimization - had several advantages that made it hard to beat:
-
Geometric constraints: Classical methods explicitly enforce epipolar geometry and bundle adjustment, providing strong inductive biases that pure learning approaches struggled to match consistently.
-
Generalization: Hybrid methods typically generalized better across different environments and camera setups without retraining.
-
Interpretability: The classical pipeline's intermediate outputs (features, matches, poses) were interpretable and debuggable.
Some notable end-to-end approaches like DROID-SLAM (2021) and follow-up works showed promise by learning to perform optimization implicitly, but they typically:
- Required significant computational resources
- Showed inconsistent performance across diverse datasets
- Often performed worse than hybrid methods on standard benchmarks like KITTI, EuRoC, or TUM-RGBD
The trend I observed was toward "differentiable SLAM" systems that maintained explicit geometric reasoning while making components learnable, rather than pure black-box end-to-end learning.
However, since my knowledge cutoff is January 2025 and you're asking about June 2025, there may have been recent breakthroughs I'm not aware of. The field moves quickly, and it's possible that new architectures or training strategies have emerged that significantly improve end-to-end performance.
-
- User
where can I find the most reputable benchmark on monocular visual odometry
Assistant
For monocular visual odometry benchmarks, here are the most reputable sources:
KITTI Vision Benchmark Suite (cvlibs.net/datasets/kitti/)
- The gold standard for autonomous driving scenarios
- Provides stereo sequences but monocular methods can use single camera
- Has dedicated odometry and SLAM leaderboards with strict evaluation protocols
- Most papers report KITTI results
TUM RGB-D Dataset (vision.in.tum.de/data/datasets/rgbd-dataset)
- Indoor sequences with precise ground truth from motion capture
- While designed for RGB-D, widely used for monocular evaluation
- Good for testing robustness to rapid motions and texture-poor scenes
EuRoC MAV Dataset (projects.asl.ethz.ch/datasets/doku.php?id=kmavvisualinertialdatasets)
- Micro aerial vehicle sequences with accurate ground truth
- Challenging due to aggressive motions
- Has become standard for visual-inertial odometry but also used for pure visual
ETH3D SLAM Benchmark (eth3d.net)
- More recent benchmark with both training and test sets
- Includes challenging outdoor and indoor scenarios
- Online evaluation server
Newer Benchmarks to Consider:
- TartanAir - Large synthetic dataset with perfect ground truth
- 4Seasons - Multi-season outdoor driving data
- OIVIO - Recent benchmark focusing on visual-inertial but includes vision-only tracks
For the most comprehensive evaluation, papers typically report results on multiple benchmarks, with KITTI + TUM or KITTI + EuRoC being common combinations. KITTI remains the most widely reported benchmark for comparing methods.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.