Fusing Pseudo Labels with Weak Supervision for Dynamic Traffic Scenarios

TL;DR

This paper presents a weakly-supervised label unification approach that combines pseudo labels from multiple models to improve object detection in diverse and challenging traffic scenarios, enhancing ADAS robustness.

Contribution

It introduces a novel pipeline that unifies labels from heterogeneous datasets, rectifies bias, and retrains a resilient detection model for dynamic traffic environments.

Findings

Significant improvement in object detection accuracy across diverse datasets.

Enhanced model robustness against domain shifts and challenging conditions.

Effective unification of heterogeneous labels for better generalization.

Abstract

Advanced Driver Assistance Systems (ADAS) have made significant strides, capitalizing on computer vision to enhance perception and decision-making capabilities. Nonetheless, the adaptation of these systems to diverse traffic scenarios poses challenges due to shifts in data distribution stemming from factors such as location, weather, and road infrastructure. To tackle this, we introduce a weakly-supervised label unification pipeline that amalgamates pseudo labels from a multitude of object detection models trained on heterogeneous datasets. Our pipeline engenders a unified label space through the amalgamation of labels from disparate datasets, rectifying bias and enhancing generalization. We fine-tune multiple object detection models on individual datasets, subsequently crafting a unified dataset featuring pseudo labels, meticulously validated for precision. Following this, we retrain a solitary object detection model using the merged label space, culminating in a resilient model proficient in dynamic traffic scenarios. We put forth a comprehensive evaluation of our approach, employing diverse datasets originating from varied Asian countries, effectively demonstrating its efficacy in challenging road conditions. Notably, our method yields substantial enhancements in object detection performance, culminating in a model with heightened resistance against domain shifts.