Abstract
DistanceNet is a novel Recurrent Convolutional Neural Network (RCNN) designed to solve the "scale ambiguity" problem in monocular visual odometry. By analyzing sequences of images, it predicts the absolute traveled distance with high precision, outperforming existing deep learning pose estimators and classical monocular SLAM methods on the KITTI dataset.
Introduction: The Challenge of Scale Ambiguity
In computer vision, "scale ambiguity" remains a fundamental hurdle for monocular (single-camera) systems. A single camera cannot inherently distinguish between a large object moving quickly at a distance and a small object moving slowly nearby. Traditional monocular Visual SLAM (vSLAM) and Visual Odometry (VO) methods can track relative motion but typically require external sensors—such as speedometers or IMUs—to determine absolute real-world scale.
While some deep learning approaches attempt to resolve this by estimating per-frame depth maps, they often neglect the temporal flow of information across multiple frames. DistanceNet bridges this gap by learning to estimate absolute distance directly from the temporal dynamics of an image sequence.
Methodology: The DistanceNet Architecture
The authors propose an end-to-end many-to-one RCNN architecture. The pipeline consists of three primary stages:
- Feature Extraction (CNN): The model utilizes a modified FlowNetC architecture. It processes pairs of consecutive images (resized to 768 x 256) to extract geometric features representing visual motion.
- Temporal Learning (RNN): These features are fed into a Recurrent Neural Network comprising two Bidirectional LSTM (Bi-LSTM) layers. By analyzing a window of 10 frames (9 pairs), the RNN captures the vehicle's movement dynamics over time, including momentum and acceleration.
- Ordinal Regression Output: Instead of standard regression, the model employs ordinal regression. It outputs a "multi-hot encoded" vector where bits represent incremental distance thresholds (e.g., "at least 1m," "at least 2m"). This approach exploits the natural mathematical order of distances, leading to more stable training.
The network is trained using Focal Loss, which prioritizes "hard" examples—distances the model frequently mispredicts—preventing overconfidence on common, easier cases.
Experimental Setup
DistanceNet was evaluated using the KITTI Odometry Benchmark. The researchers utilized various combinations of the 22 sequences for training and testing. Notably, Sequence 01 (highway) was excluded from training due to its high-speed environment, which differed significantly from the urban and rural settings of other sequences, posing a challenge for generalization with limited data. The model was trained against ground-truth pose-annotated data where absolute traveled distance is known.
Key Results and Performance
DistanceNet demonstrated superior performance in predicting traveled distance compared to both classical and modern deep learning methods.
- Accuracy: On KITTI sequences 03-07 and 10, DistanceNet achieved a Root Mean Squared Error (RMSE) of 0.3901 meters, significantly outperforming the state-of-the-art SRNN-channel (0.5033) and the classical PMO method (0.7463).
- Consistency: The model achieved an AccDev (Accuracy within 1-meter deviation) of 99.16%, indicating that predictions remain within a tight, usable margin of error.
Ablation studies confirmed that the specific architectural choices—using FlowNetC over FlowNetS and ordinal regression over standard MSE regression—were critical to achieving these results.
Conclusion and Future Work
DistanceNet effectively uses distance as a proxy for scale. Since the camera frame rate is constant (10 fps in KITTI), estimating distance between frames is mathematically equivalent to estimating absolute speed, providing the necessary scale factor for monocular SLAM.
Currently, DistanceNet focuses on translation (distance) rather than full 6-Degree-of-Freedom (6-DoF) pose estimation. Future work involves integrating DistanceNet as a "virtual speedometer" module within full monocular SLAM systems, where classical algorithms handle geometry and rotation while DistanceNet resolves scale drift in real-time. Additionally, expanding training data to include diverse speed regimes will improve the model's generalizability.
Summary updated