OmniTraffic 3D Environments

OmniTraffic Blender Environments

3D traffic scene environments for Blender, compatible with the OmniTraffic pipeline.

Overview

This package contains 12 reconstructed real-world intersections as Blender .blend files. These environments are used by the Scenario Collector to render synthetic traffic images for the OmniTraffic VQA benchmark.

Usage

1. Download and extract this archive
2. Copy the traffic_scenarios/ folder into the OmniTraffic repository root
3. The .blend files will be located at traffic_scenarios/<location>/env.blend

Alternatively, place each env.blend into its corresponding scenario directory:

OmniTraffic/traffic_scenarios/<location>/env.blend

Rendering

Use the provided scripts in scenario_collector/renderers/ to render images:

# Single scene
./scenario_collector/renderers/render_single.sh --scenario <path> --start 0 --end 600

# Batch rendering
./scenario_collector/renderers/render_batch.sh --scenario <path1> --scenario <path2>

Scenarios

Location	Description
Beijing_Beihuan	Urban intersection in Beijing
Beijing_Beishahe	Residential area intersection
Beijing_Changjianglu	Commercial district intersection
Beijing_Gaojiaoyuan	Suburban intersection
Beijing_Pinganli	Central urban intersection
Beijing_Yongrunlu	High-traffic corridor
Chengdu_Chenghannanlu	Chengdu downtown intersection
Chengdu_Guanghua	Technology district intersection
France_Massy	European urban intersection
Hongkong_YMT	Hong Kong urban intersection
SouthKorea_Songdo	Korean smart city intersection
Tianjin_zhijingdao	Tianjin main road intersection

File Structure

traffic_scenarios/
├── Beijing_Beihuan/env.blend
├── Beijing_Beishahe/env.blend
├── Beijing_Changjianglu/env.blend
├── Beijing_Gaojiaoyuan/env.blend
├── Beijing_Pinganli/env.blend
├── Beijing_Yongrunlu/env.blend
├── Chengdu_Chenghannanlu/env.blend
├── Chengdu_Guanghua/env.blend
├── France_Massy/env.blend
├── Hongkong_YMT/env.blend
├── SouthKorea_Songdo/env.blend
└── Tianjin_zhijingdao/env.blend

OmniTraffic Human Validation Data

OmniTraffic Human Validation Dataset

This dataset is a comprehensive benchmark for traffic scene visual question answering. It contains 100 carefully selected questions to evaluate human performance on traffic scene understanding tasks.

Dataset Contents

• Synthetic/Simulation Images (qa_dataset): Generated from traffic simulation software

  • Multiple cities: Beijing, Chengdu, Tianjin, Seoul, etc.
  • Various traffic scenarios: normal traffic, accidents, congestion, etc.

• Real-world Images (qa_dataset_real): Real traffic scenes

• Question Types:

  • Vehicle Analysis (counting, classification)
  • Road Infrastructure (lane detection)
  • Special Vehicles (police cars, ambulances, fire trucks)
  • Special Events (traffic accidents, obstructions)

  Usage

  1. Download and extract image.zip
  2. Navigate to human_validation/web_app
  3. Run uv run python app.py

  📊 Use together with the OmniTraffic benchmark to evaluate model performance against human baselines