Abstract
This work is devoted to the traffic sign detection on images using deep learning methods. We focus on the problem of detector transfer to new datasets with different road signs. We present an algorithm for distilling a set of unlabelled data to select the most informative images to be labeled. This method allows to significantly reduce the amount of data labeling with a small decline of detector performance.
Introduction
Deep learning has revolutionized object detection, but it requires vast amounts of labeled data. For autonomous driving, a traffic sign detector trained in one region (e.g., Russia) often performs poorly in another (e.g., China) due to variations in sign design, lighting, and environmental conditions. Traditionally, this necessitates collecting and manually labeling thousands of new images for every new region—a slow and expensive process. This paper investigates "data distillation": identifying a subset of the most informative images for labeling to maintain high performance while minimizing human effort.
Methodology
The authors explore strategies for selecting training data, moving from simple baselines to a novel complexity-based approach for unlabeled data.
Complexity Prediction
To avoid the need for initial labels in the target domain, the authors propose training a "complexity predictor." This is a small convolutional neural network based on the Xception architecture. It takes a 300x300 image and outputs a scalar representing the predicted complexity (or "hardness") for a detector.
DupletLoss
Rather than regressing an exact error value, which can be highly noisy, the authors introduced DupletLoss. This loss function trains the network to correctly rank pairs of images by their relative complexity. By focusing on ranking rather than absolute value, the model more effectively identifies the most informative samples for finetuning.
Experimental Setup and Results
The study utilized the RTSD (Russian Traffic Sign Dataset) as the source domain and the TTK100 (Tsinghua-Tencent 100K) Chinese dataset for finetuning. Performance was measured using Area Under the Curve (AUC) across various sampling rates.
Key findings include:
- Efficiency: Using loss-ranking to select only half of the target dataset yielded an average AUC of 0.8873, nearly matching the performance of the full dataset.
- Complexity as a Proxy for Rarity: High-loss images often contain rare sign classes. Selecting these "complex" images naturally balances the dataset, focusing the model on edge cases it hasn't yet mastered.
- Superiority over Random Sampling: Complexity-informed selection consistently outperformed random sampling, particularly at lower sampling rates (e.g., 1/16).
Conclusion
Data distillation via complexity prediction offers a viable path to reducing labeling costs in traffic sign detection. While performance drops at extreme compression levels (1/32) due to the selection of noisy or low-quality frames, the ranking-based approach effectively identifies informative samples. Future work will focus on distinguishing between "useful complexity" (rare signs) and "useless complexity" (sensor noise).
Summary updated