BRSSD10k : A SEGMENTATION DATASET \\OF BANGLADESHI ROAD SCENARIO

Mirza Nihal Baig, Mahdi Murshed Patwary, Husne Ara Chowdhury, Md. Shahidur Rahman

Keywords: Instance Segmentation, Computer Vision, Dataset, Autonomous Driving, Bangadeshi Road

TL;DR: BRSSD10k is Bangladesh's first instance segmentation dataset for autonomous driving, with 10,082 high-res images from 9 cities. It offers detailed road element annotations, providing a key benchmark for AI models in diverse South Asian road scenarios

Abstract: In this paper, we present a novel Bangladeshi Road Scenario Segmentation Dataset designed to advance autonomous driving technologies under the challenging and diverse road conditions of Bangladesh. This comprehensive instance segmentation dataset comprised 10,082 high-resolution images captured across nine major cities, including Dhaka, Sylhet, Chittagong, and Rajshahi, addressing the critical need for region-specific computer vision data in developing countries. Unlike existing autonomous driving datasets that primarily focus on western road conditions, BRSSD10k encompasses a wide range of environments unique to Bangladesh, including unstructured urban areas, hilly terrains, village roads, and densely populated city centers. The dataset features instance segmentation annotations with classes specifically tailored to reflect the distinctive elements of Bangladeshi roads, such as rickshaws, CNGs (auto-rickshaws), informal roadside stalls, and various nonstandard vehicles. To demonstrate its utility as a benchmarking tool for autonomous driving systems, we present comparative results from several state-of-the-art instance segmentation models tested on this dataset, achieving an mAP of 0.441. This evaluation not only showcases the dataset's effectiveness in assessing model performance but also underscores the need for adaptive algorithms capable of handling diverse and unpredictable urban environments in the context of autonomous navigation.

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9147