The authors of IITM-HeTra: Dataset for Vehicle Detection in Heterogeneous Traffic Scenarios emphasize the utility of video image processing from traffic camera feeds for tasks like counting and classifying vehicles, estimating queue length, determining traffic speed, and tracking individual vehicles. In contrast to homogeneous traffic, heterogeneous traffic involves various vehicle types that do not adhere to lane discipline, making vehicle detection particularly challenging, especially when vehicles are occluded, a common occurrence in such scenarios.

Recent advancements in Deep Learning have demonstrated significant potential in addressing various computer vision tasks, including object recognition, detection, and tracking. However, it’s worth noting that training deep learning models necessitates extensive labeled datasets, which are both time-consuming and expensive to obtain. To address this challenge, the authors propose a solution involving data augmentation. Specifically, they augment an existing large, general (non-traffic) dataset with a small, low-resolution dataset of heterogeneous traffic (collected by the authors themselves). This augmentation approach yields state-of-the-art vehicle detection performance. It’s important to highlight that, to the best of the authors’ knowledge, the dataset they collected, known as IITM-HeTra, represents the first publicly available labeled dataset for heterogeneous traffic.

To ensure that data are temporally uncorrelated, the authors of the study sampled a frame every two seconds from multiple video streams. A total of 2400 frames were extracted. 2400 frames were manually labeled under different vehicle categories by the authors. After careful scrutiny and elimination of unclear images, the number of available frames was reduced to 1417. The dataset was then divided by the authors into a trainval set (1202 images) and a test set (216 images), and this split was retained for all experiments. Initially, eight different vehicle classes commonly seen in Indian traffic were defined by the authors. Some of these classes were similar, while the two classes had fewer labeled instances; these were merged into similar-looking classes by the authors. For instance, in the dataset, there were different categories for small cars, SUVs, and sedans, which were merged under the “light motor vehicle (LMV)” category by the authors. The collected dataset contained a total of 6319 labeled vehicles. This included 3294 two-wheelers, 279 heavy motor vehicles (HMV), 2148 cars, and 598 autorickshaws. A second dataset was created by the authors by merging cars and autorickshaws together into the “light motor vehicle (LMV)” class. Approximately 25.2% of the vehicles were occluded, according to the authors.

Please note, that the original light motor vehicle (LMV) and cars classes are merged into a single cars class. heavy motor vehicles (HMV) class is renamed into bus, and two-wheelers is renamed to person respectfully.