Deepa Mane
Sandhya Arora2- 1Department of Computer Engineering, Smt. Kashibai Navale College of Engineering, SPPU, Pune, India
- 2Department of Computer Engineering, Cummins College of Engineering for Women, SPPU, Pune, India
- 3Department of Information Technology, Pune Institute of Computer Technology, SPPU, Pune, India
1 Introduction
Autonomous driving systems are inherently reliant on strong object detection to effectively sense and traverse intricate real-world scenes. While deep learning, and more specifically convolutional neural networks (CNNs), have powered much of the growth in this area, most public benchmarks like KITTI (Varma et al., 2019), BDD10K and Cityscape (Maddern et al., 2017) draw from strongly structured, rule-abiding traffic scenarios found in Western nations. These settings are quite far from those in countries such as India, where unstructured traffic, uneven infrastructure, and extreme environmental fluctuations prevail. Because diversity is so high in India, it provides an ideal setting to create generalizable perception models able to deal with uncertainty and real-world complexity. Some Indian datasets like the Indian Driving Dataset (IDD) (Varma et al., 2019; Paranjape and Naik, 2022), ITD (Agarwal et al., 2024), DATS (Srinath et al., 2020), NITCAD (Krizhevsky et al., 2017), and the latest multimodal TIAND dataset (Yang et al., 2019) Sdac (Gong et al., 2024) have been proposed to overcome these challenges. Although all of them provide useful insights, they generally concentrate on individual tasks (e.g., semantic segmentation), few object classes, or mainly urban environments, leaving a gap for an extensive, large-scale dataset focused on heterogeneous object detection in unstructured Indian traffic scenes. IndiaScene365: A transfer Learning dataset for Indian Scene Understanding in diverse weather conditions comprises images from traffic environments that truly represents Indian roads. Unlike other standard available datasets that feature well-organized traffic scenes from various global locations, this collection includes vehicle types such as motorcycles, auto-rickshaws, and animal-drawn carts, which are characteristic of Indian road conditions but absent from worldwide datasets. Researchers interested in using this dataset can be downloaded directly from Mendeley data at https://data.mendeley.com/datasets/pwffhg6nhz/1 or contact the author directly (Mane, 2022).
2 Background
In recent years, numerous datasets for advanced driver assistance systems and autonomous car driving have emerged, primarily focusing on structured and well-defined driving environment. These typically feature well-defined lanes, a limited number of clearly categorizable objects, minimum difference between object and background and strict adherence to traffic rules. The Oxford Robotcar dataset pioneered large-scale real-world data collection (Maddern et al., 2017), including significant representation of challenging visual traffic scenes such as low illumination, rain, nighttime and foggy but it might not cover extreme weather conditions like snow, heavy rain, or fog in sufficient detail (Yin et al., 2024). This limits the ability to train models that can generalize well to such conditions. Foggy Zurich (Sakaridis et al., 2018) and Foggy Cityscapes (Paranjape and Naik, 2022) introduced synthetic fog images by applying masks to original images. Synthetic rain datasets (Zheng and Yoo, 2025) include Rain1400, RainyCityscapes Rain100H, and Rain12. Datasets like, BDD100K, ACDC (Dokania et al., 2023), and IDD (Varma et al., 2019; Agarwal et al., 2024), contain authentic images collected from various adverse conditions (Yin et al., 2024), including fog, rain, snow, and nighttime scenarios (Srinath et al., 2020; Krizhevsky et al., 2017). However, most of these datasets are prepared based on structured environments and have inadequate annotations for unstructured scenarios. Table 1 Comparison of major Indian and global object detection datasets shows the comparative analysis of all the other datasets we considered and novelty of our dataset. Below is the Label Hierarchy that we have used in our dataset. The dataset uses a hierarchical labeling framework based on a parent-child relationship to represent the semantic relationship between traffic entities on a broad scale and a finer-grained scale. Classes are organized into a wider parent class, which is organized into a child sub-class. For example, Vehicles → Car, Bus, Truck, Auto-rickshaw. This hierarchy supports flexible training, allowing algorithms to be trained using either flat labels for classification/segmentation for a single layer or hierarchical supervision for better generalization across similar classes.
Table 1. Comparison of major Indian and global object detection datasets.
2.1 Dataset overview
Indiascene365 is a dataset comprising over 3,000 images captured using Android smartphone (One Plus Nord CE) equipped with 50 MP super cameras and unplug car mounted stereo front and back camera. These devices produce images of quality comparable to professional cameras. The increasing prevalence of mobile phones, which are preferred over bulky cameras, makes this mobile-captured dataset more practical and representative. The images were taken across various locations in Maharashtra, India, encompassing different road types and traffic conditions as detailed in Table 2 Details of Data collection Sources. The dataset includes photos taken during the day and evening, across all seasons. The diverse locations range from bustling city streets and 6-lane national highways with varying traffic densities to rural roads with animal presence (Li et al., 2023). Table 2 showcases depicting national highways, rural roads, construction-affected highways, crowded urban streets, mountain roads with wildlife, traffic at different times of day, highways with large vehicles, and animals on roadways. The mountain roads feature sharp turns bordered by dense forests or steep cliffs and deep valleys.
Table 2. Details of data collection sources.
2.2 Data classes
The label set used in dataset preparation is identical to that in IDD (Varma et al., 2019; Agarwal et al., 2024) The hierarchy in the category of class labels adds a higher degree of complexity (Srinath et al., 2020) to our dataset compared to existing datasets like Cityscapes (Yang et al., 2019), and even when compared to adverse weather datasets such as Foggy Cityscapes (Geiger et al., 2013; Baker et al., 2011). A team of highly skilled annotators was employed to label the dataset. The labelling process for each weather condition images typically take 2–3 h, including initial annotation and review process. To ensure annotation accuracy, multiple revisions were taken, along with verifying annotations against previously defined classes and conducting a final validation by expert annotators. Indiascene365 is split into three sets corresponding to the examined conditions. We took around 1,000 lowlight and daylight, around 1,000 rain and 1,000 foggy images from selected recordings for detailed annotation process, resulting in 3,000 images depicting adverse conditions. The selection criteria focused on maximizing scene complexity and diversity (Sakaridis et al., 2021). To ensure robust evaluations, we implemented a comprehensive train-test split method with defined parameters. All images from a single drive sequence were allocated entirely to either the train or test set, guaranteeing that test data will remain novel. The test set for each weather condition included drive sequences comprising 18% to 20% of total scenes, ensuring a representative sample of various driving scenarios. In the test set, we kept the mean number of frames per drive sequence between 90% and 120% of the dataset's overall average. To ensure fairness across classes, we limited the average number of instances per image in the test set to between 80% and 120% of the entire dataset's average. To guarantee accurate pixel-level annotations, test images were required to have an average pixel ratio ranging from 0.5 to 1.4 for at least 18 classes, and from 0.4 to 1.5 for a minimum of 24 classes.
2.3 Data formats and file structure
The dataset comprises diverse road scenes, including national highways, country roads, construction zones, busy city streets, and mountain passes. It showcases traffic patterns at various times of day, as well as different weather conditions such as rain and fog. The collection features highways with large vehicles and instances of animals on roads. Mountain routes are characterized by sharp turns, with dense forests on one side and steep peaks or valleys on the other. In rural areas, trees line both sides of the roads. The aim of forming this dataset is to capture images representing various outdoor scenes throughout the city, encompassing different weather scenarios and times of day. High quality videos and images were collected using smartphone cameras. Afterward frames were extracted from video streams as images using VLC media player at 30 fps. Indiascene365 classifies objects into 34 distinct categories, organized into broader groups, as show Figure 1. A thorough dataset should include data descriptions, or annotations, alongside the images. Data annotation involves defining areas or objects within an image and creating text descriptions for them. Figure 2 represents the data structure of the data in the repository. The open-source Roboflow tool1 was used to annotate images in the proposed dataset. After completing image annotation and saving, an XML file is generated for each image and needs to maintain the different folders for different weather conditions as well as for XML and annotations files. These XML files have details about each labeled object in the image and its coordinates in the images, specifying the annotation type as “bounding box”. Additionally, .txt and .json files are generated with reference each image and stored in the respective folders. The dataset comprises diverse road scenes, including national highways, country roads, construction zones, busy city streets, and mountain passes. It showcases traffic patterns at various times of day, as well as different weather conditions such as rain and fog. The collection features highways with large vehicles and instances of animals on roads. Mountain routes are characterized by sharp turns, with dense forests on oneside and steep peaks or valleys on the other. In rural areas, trees line both sides of the roads. Figure 3 shows sample Images of diverse road conditions at different seasons and their respective annotated images with bounding boxes.” The aim of forming this dataset is to capture image representing various outdoor scenes throughout the city, encompassing different weather scenarios and times of day. High quality videos and images were collected using smartphone cameras. Afterward frames were extracted from video streams as images using VLC media player at 30 fps.
Figure 1. (a) Class distribution heatmap. (b) Weather condition distribution vs. number of images.
Figure 2. Data folder layout.
Figure 3. Sample Images of diverse road conditions at different seasons on left side and their respective annotated images with bounding boxes.
3 Materials and methods
Dataset Creation Process and Methodology: Figure 4 illustrate the steps involved in generating the dataset. Images were collected from various roadways within Pune city and its surrounding areas in Maharashtra, India. The dataset was compiled by the authors and team members using a standard Android smartphone camera and car mounted stereo front and back camera over a 1-year period. Consequently, the image count is lower compared to other datasets. However, image collection will be used for semantic segmentation soon, and the dataset will be revised periodically.
Figure 4. Dataset generation process.
3.1 Experiments
3.1.1 . Experiment 1
A Faster RCNN-based object detection model was pre-trained on global datasets (Cityscapes, BDD100K) and fine-tuned on IndiaScene365 to evaluate cross-domain adaptation performance.
3.1.2 Experiment 2
A YOLOv11-based object detection model was pre-trained on global datasets (Cityscapes, BDD100K) and fine-tuned on IndiaScene365 to evaluate cross-domain adaptation performance.
3.2 Classification results
To evaluate the effectiveness of our dataset in developing accurate deep learning models for classifying the objects in scene understanding we evaluated with object detectors like Faster RCNN and a Yolov11. More details about the methodology and configuration of the experiments, are described in 3.1. Here we present the results obtained for both the experiments. Table 3 summarizes the results and Figure 5 Shows the confusion matrix for the second experiment.
Table 3. Classification results.
Figure 5. Confusion matrix.
3.3 Data quality assurance
To confirm the reliability and diversity of the dataset, we conducted an image quality and redundancy assessment. Quality was assessed with objective metrics, including BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator) for perceptual quality, variance of the Laplacian level as a measure of sharpness, and mean brightness and contrast measures to evaluate light consistency. Frames that exhibited extremely low levels of sharpness or presented abnormal light conditions were either tagged for manual review or discarded. Redundancy arising from sequential video input required an approach to identify near-duplicate frames through perceptual hash (pHash)-based analysis of visual similarity. Selecting a cutoff of 95% similarity, we disseminated frames that responded to the threshold and were visually distinct. In this way, we took slight actions to reduce redundancy and maintain level of representational diversity across different weather and scenes, limiting any implications of redundancy or similarity for the representational sample.
4 Limitations and future scope
This transfer learning dataset is obtained from sparse domains, which could limit its generalizability and versatility in low-resource settings. Despite the variety provided by Indiascene365, there are still some limitations, providing openings for future improvement:
Geographically this dataset largely reflects western Indian regions and fails to completely cover the varied terrains, traffic conditions, and signage observed in other parts of India—e.g., hilly regions or densely populated urban areas in the north and northeast. Safety-critical but less frequent classes (e.g., ambulances, police cars) are not well-represented, which results in difficulties for robust model training on long-tail classes. Indiascene365 has just RGB images at present. Adding other modalities like LiDAR would further improve perception performance in the presence of adverse weather and complex road geometry.
Limited Annotation Depth: The data offers only 2D bounding box annotations. Including semantic segmentation masks or 3D bounding boxes (e.g., point-level LiDAR) would make it more useful for more general autonomous vehicle perception tasks.
looking forward Indiascene365 as a starting point benchmark to propel research in strong, multimodal perception for unstructured roadway environments.
Data availability statement
The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/supplementary material.
Ethics statement
Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.
Author contributions
DM: Conceptualization, Data curation, Formal analysis, Methodology, Resources, Software, Writing – original draft. SA: Formal analysis, Investigation, Project administration, Supervision, Writing – review & editing. SS: Resources, Software, Validation, Visualization, Writing – review & editing.
Funding
The author(s) declared that financial support was not received for this work and/or its publication.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher's note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Footnotes
References
Agarwal, A., Thombre, A., Kedia, K., and Ghosh, I. (2024). “Itd: Indian traffic dataset for intelligent transportation systems,” in 2024 16th International Conference on COMmunication Systems and NETworkS (COMSNETS) (Bengaluru: IEEE), 842–850. doi: 10.1109/COMSNETS59351.2024.10427394
Baker, S., Scharstein, D., Lewis, J. P., Roth, S., Black, M. J., and Szeliski, R. (2011). A database and evaluation methodology for optical flow. Int. J. Comput. Vis. 92, 1–31. doi: 10.1007/s11263-010-0390-2
Dokania, S., Hafez, A. H. A., Subramanian, A., Chandraker, M., and Jawahar, C. V. (2023). “Idd-3d: Indian driving dataset for 3d unstructured road scenes,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 4482–4491. Available online at: https://cdn.iiit.ac.in/cdn/cvit.iiit.ac.in/images/ConferencePapers/2023/IDD-3D.pdf (Accessed October 23, 2025).
Geiger, A., Lenz, P., Stiller, C., and Urtasun, R. (2013). Vision meets robotics: the kitti dataset. Int. J. Robot. Res. 32, 1231–1237. doi: 10.1177/0278364913491297
Gong, L., Zhang, Y., Xia, Y., Zhang, Y., and Ji, J. (2024). Sdac: a multimodal synthetic dataset for anomaly and corner case detection in autonomous driving. Proc. AAAI Conf. Artif. Intell. 38, 1914–1922. doi: 10.1609/aaai.v38i3.27961
Krizhevsky, A., Sutskever, I., and Hinton G, E. (2017). ImageNet classification with deep convolutional neural networks, Commun. ACM 60, 84–90. doi: 10.1145/3065386
Li, J., Kaur, S. S., and Do, M. N. (2023). “Domain adaptive object detection for autonomous driving under foggy weather,” in Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV) (Waikoloa, HI: IEEE), 4332–4341. doi: 10.1109/WACV56688.2023.00068
Maddern, W., Pascoe, G., Linegar, C., and Newman, P. (2017). 1 year, 1000 km: the oxford robotcar dataset. Int. J. Robot. Res. 36, 3–15. doi: 10.1177/0278364916679498
Paranjape, B. A., and Naik, A. A. (2022). Dats_2022: a versatile indian dataset for object detection in unstructured traffic conditions. Data Brief 43:108470. doi: 10.1016/j.dib.2022.108470
Sakaridis, C., Dai, D., and Van Gool, L. (2018). Semantic foggy scene understanding with synthetic data. Int. J. Comput. Vis. 126, 973–992. doi: 10.1007/s11263-018-1072-8
Sakaridis, C., Wang, H., Li, K., Zurbrugg, R., Jadon, A., Abbeloos, W., et al. (2021). ACDC: The Adverse Conditions Dataset with Correspondences for Robust Semantic Driving Scene Perception (IEEE). doi: 10.1109/ICCV48922.2021.01059
Srinath, N. G. S. S., Joseph, A. Z., Umamaheswaran, S., Priyanka, C. L., Nair, M., and Sankaran, P. (2020). NITCAD-Developing an object detection, classification and stereo vision dataset for autonomous navigation in Indian roads. Proc. Comput. Sci. 171, 207–216. doi: 10.1016/j.procs.2020.04.022
Varma, G., Subramanian, A., Namboodiri, A., Chandraker, M., and Jawahar, C. V. (2019). “IDD: a dataset for exploring problems of autonomous navigation in unconstrained environments,” in 2019 IEEE Winter Conference on Applications of Computer Vision (WACV) (IIT Hyderabad, IEEE), 1743–1751. doi: 10.1109/WACV.2019.00190
Yang, W., Tan, R. T., Feng, J., Guo, Z., Yan, S., and Liu, J. (2019). Joint rain detection and removal from a single image with contextualized deep networks. IEEE Trans. Pattern Anal. Mach. Intell. 42, 1377–1393. doi: 10.1109/TPAMI.2019.2895793
Yin, H., Wang, P., Liu, B., and Yan, J. (2024). An uncertainty-aware domain adaptive semantic segmentation framework. Auton. Intell. Syst. 4:15. doi: 10.1007/s43684-024-00070-0
Keywords: deep learning, pre-trained models, scene understanding, transfer learning, domain adaptation
Citation: Mane D, Arora S and Shelke S (2026) IndiaScene365: a transfer learning dataset for Indian scene understanding in diverse weather condition. Front. Artif. Intell. 8:1669512. doi: 10.3389/frai.2025.1669512
Received: 19 July 2025; Revised: 21 October 2025;
Accepted: 12 December 2025; Published: 09 January 2026.
Edited by:
Mohammed El-Abd, American University of Kuwait, KuwaitReviewed by:
Ruben Cornelius Siagian, State University of Medan, IndonesiaCopyright © 2026 Mane, Arora and Shelke. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Deepa Mane, ZGVlcGFibWFuZUBnbWFpbC5jb20=