Enhancing Odometry Models for Robustness through Explainability-Driven Data Augmentation

Sammanfattning

Visual Odometry (VO) is essential for autonomous navigation in GPS-denied environments. Although deep learning-based VO models have demonstrated strong performance in complex scenarios, their black-box nature makes it difficult to interpret predictions or diagnose failures. This lack of transparency limits their reliability in challenging conditions, such as motion blur, occlusions, dynamic elements, and weak feature regions. This thesis explores the use of explainability techniques to enhance the robustness and interpretability of monocular deep learning-based VO systems. Grad-CAM and Smooth Grad-CAM++ were applied to a pre-trained TSformer-VO model evaluated on the KITTI dataset. High-error frames were identified using Absolute Trajectory Error (ATE), and attention heatmaps were correlated with semantic segmentation to expose failure patterns and their spatial context. Failure cases were categorized based on semantic attention characteristics, revealing that 86% of the model’s errors stemmed from motion blur or weak features, which caused attention to drift toward irrelevant background regions. The remaining 14% of cases involved semantically reasonable focus but still produced significant pose errors, likely due to domain shifts or scene ambiguity. Based on these insights, targeted data augmentation techniques—including motion blur simulation and contrast/lighting adjustments—were designed and applied to the most failure-prone frames. Retraining the TSformer-VO model on the augmented dataset led to improvements in pose estimation accuracy, as evidenced by reductions in ATE, and clearer, more semantically aligned attention maps. These results confirm that explainability methods are not only valuable for diagnosing model behavior but can also inform effective interventions to improve model performance and resilience in real-world conditions.