back to work
03Applied AI Research·Research

VLM Transportation Research

A research project evaluating Vision-Language Models on Indian road-scene understanding across paired dashcam and top-view data.

role
Dataset engineering, annotation schema, evaluation
stack
Python · Qwen2.5-VL · Pandas · Google Colab · Custom Eval Harness

Overview

A research effort probing how well modern Vision-Language Models understand Indian road scenes when given paired viewpoints, structured prompts, and consistent evaluation.

Problem

VLMs perform impressively on generic benchmarks, but their behavior on unstructured, non-Western road scenes is under-explored. This project builds the dataset and harness needed to ask that question honestly.

Approach

  1. 01Pair dashcam and top-view footage; match frames to GPS timestamps.
  2. 02Design a compact annotation schema covering actors, hazards, and scene state.
  3. 03Prepare a curated evaluation subset with reviewer-verified labels.
  4. 04Run zero-shot VLM evaluation with a fixed prompt template.
  5. 05Compare multimodal inputs against single-view baselines with statistical and qualitative analysis.

My contributions

  • Dataset engineering: collection, cleaning, frame-GPS matching.
  • Annotation schema design and labeling guidelines.
  • Evaluation pipeline and reporting.
  • Qualitative failure-mode analysis across scene categories.

Outcomes

  • Reusable dataset structure for future VLM road-scene work.
  • A clear picture of where zero-shot VLMs help and where they break on Indian road data.
// scope · what this is not
  • × No publication, acceptance, or peer review claimed.
  • × Results are internal and exploratory.
// next.case_studies