Answers are grounded in this paper and cite their sources.
Deep learning has been recently used to detect roadside attributes with some success. However, deep learning models require many samples to be annotated for each attribute to produce acceptable accuracy for the industry. It is very challenging to annotate complex roadside data with many attributes accurately. Therefore, in this paper, we propose a new technique that can learn from small training data. The novel idea is to develop a dedicated deep learner for each class using a one-vs-all classification strategy, where each learner is trained to distinguish a specific target class from all remaining classes. Models were trained over 40,000 iterations, with selections made at specific intervals based on the highest classification accuracy and lowest misclassification rate, followed by a fusion technique to combine the best-trained models. In addition to the proposed technique, we have also created small image datasets for each attribute from the video data provided by our industry partner. The images in the datasets cover over 61 road safety attributes. Each image in the benchmark is annotated with a separate Red, Green, and Blue (RGB) color. We have evaluated the proposed technique on a real-world dataset and on a custom test dataset prepared by manually annotated Ground Truth (GT) data. The results for many attributes showed that the proposed technique can learn from a small number of samples and achieve 100% attribute wise accuracy and above 90% pixel wise accuracy in less than 40, 000 iterations. Most of the attributes exhibited higher attribute wise accuracy compared to existing techniques such as DeepLab version 3 + (DeepLabV3 +) and Fully Convolutional Network with 8 (FCN8) techniques.
No takes yet. Share an insight, caveat, or question.
Sanjeewani et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: