BhopalMLS — Dense Urban Mobile LiDAR Dataset

Video

A three-minute overview: the Azad Market scenes in the dataset viewer, the 12-, 21- and 68-class taxonomies and the instance labels, the benchmark, and the cross-region transfer finding. The viewer scenes are subsampled for display; the dataset is denser. The narration, in English or Hindi, is synthesised.

Overview

BhopalMLS captures narrow commercial corridors, mixed traffic, overhead wiring, roadside objects, informal shopfronts, and dense vegetation from handheld mobile laser scanning. Collected in an unplanned and heterogeneous dense city setting in India, it challenges state-of-the-art models trained on well-planned western datasets.

The dataset is organized as a 68-class V10 taxonomy, a 12-class coarse semantic benchmark, and a 21-class fine-grained benchmark for object-heavy scene understanding, highlighting extreme geometric clutter and diverse region-specific categories.

Annotation Samples

Labelled scenes beside the 360-degree imagery they were coloured from, the semantic labels beside the raw LiDAR intensity, and wider street scenes with zoomed details of small and rare classes. Faces in the photos are blurred.

Interactive Scene Explorer

Explore partial 3D point clouds from the dataset directly in your browser. Please note that these scenes are provided solely for visualization and understanding; certain classes, such as road surfaces, have been partially sampled to optimize browser performance. Select a scene below and use the toolbar to toggle between raw RGB, semantic taxonomies, and instance segmentation modes. Click on objects with the 'Interact' tool to inspect them.

Loading

Instance Gallery

One annotated instance per fine class, next to the street-view crop it was annotated from. Each crop is a 90° cut of the 360° panorama, framed on the instance. The 3D view opens from the scanned side and can be turned, zoomed and recoloured by intensity, height or class. Faces and number plates are blurred.

Benchmarking Foundation Models

BhopalMLS benchmarks state-of-the-art supervised models and pretrained 3D foundation models. Fine-tuned foundation models successfully transfer to complex, unstructured scenes, significantly outperforming models trained from scratch.

Supervised vs Foundation

Fine-tuned foundation models outperform supervised training from scratch: Utonia(L) reaches 71.28% mIoU against PTv3's 68.04%, and 72.12%, the best in Table 4, when trained jointly with the 6 moving subclasses (†).

PTv3 (Scratch): 68.04% mIoU
Utonia(L) Fine-tuned: 71.28% mIoU
Utonia(L)† Fine-tuned (best): 72.12% mIoU

Asymmetric Transferability

Cross-region experiments (Ke-MLS, Toronto-3D, Paris-Lille-3D, SemanticKITTI) reveal that models trained on BhopalMLS generalize to other benchmarks better than the reverse on every pair (by up to 36 mIoU points). Controlled retraining traces this to the coverage of BhopalMLS's subtype inventory, not to dataset size or corridor geometry.

BhopalMLS → Target: Strong Transfer
Target → BhopalMLS: Severe Drop

Per-class IoU heatmaps for the 12-class and 21-class semantic segmentation benchmarks — Per-class IoU (%) for 12 classes (A) and 21 classes (B): supervised 9ch baselines, fine-tuned foundation models, the † joint moving-class variants and, for 12 classes, Utonia(L) C2 (no Lovász); best per column outlined, mIoU at the right. Supplementary Fig. 8.

Row-normalised 12-class confusion matrices of Sonata, Concerto (O), Concerto (L) and Utonia fine-tuned — Row-normalised confusion matrices (12 classes) of the four fine-tuned foundation models on the test set (486 scenes); diagonal = recall, red dashed = confusions of 10% or more. Supplementary Fig. 13.

Taxonomy & Class Statistics

Explore the hierarchical label structure and dataset-level point counts. The taxonomy defines static and dynamic (moving) classes across the coarse (12-class) and fine-grained (21-class) benchmarks, the 62 static classes of the primary benchmark, and the full 68-class taxonomy.

Instance Segmentation Benchmarks

BhopalMLS provides rich instance-level annotations for all "thing" classes. Below are the quantitative baseline results and qualitative error maps for the 12-class and 21-class instance segmentation tracks.

12-class instance setting

Stuff: Ground, Vegetation, Road, Wire.
Things: Building, Pole and Tower, Wire-Structure Connector, Vehicle, Board, Fencing, Human and Animal, Road Side Object.

Best Utonia(L) FT mAP50: 44.07%
Best Utonia(L) FT mAP: 33.09%

21-class instance setting

Stuff: Ground, Vegetation, Road, Wire.
Things: 17 fine classes including Stairs, 2/3/4-Wheeler, Shop Shade, Mannequin, and Road Side Object (other).

Best Utonia(L) FT mAP50: 39.83%
Best Utonia(L) FT mAP: 30.63%

BhopalMLS 12-class qualitative instance segmentation predictions — 12-class instance predictions and error maps.

BhopalMLS 21-class qualitative instance segmentation predictions — 21-class instance predictions and error maps.

Citation

If you use BhopalMLS in your research, please cite our paper:

@inproceedings{chauhan2027bhopalmls,
  title     = {BhopalMLS: A LiDAR Dataset for Semantic Scene Understanding in Heterogeneous Urban Environments},
  author    = {Chauhan, Pyare Lal and Kumar, Vaibhav and Lohani, Bharat},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}