Lê Chí Đại
← Projects
Computer vision · Graphics · 2026

VisionDrive3D

A hierarchical rendering engine for generating synthetic datasets that mimic multi-camera autonomous driving configurations.

RoleCo-engineer
Year2026
ToolsPython · OpenGL · PyTorch

Project visuals

Figures from project reports, pitch decks, and saved experiment outputs. Select an image to open it at full resolution.

A rendered street scene with buildings, traffic lights, and cars.
RGB output from synthetic scene 00049. Source ↗
The same street scene rendered as a color-coded depth map.
Depth visualization of the same scene; the stored ground truth is metric depth. Source ↗
Individual cars and scene objects distinguished by colored instance masks.
Instance segmentation output generated alongside the RGB image. Source ↗
Projected blue 3D bounding boxes around cars in the rendered scene.
Projected 3D bounding-box annotations. Source ↗
Two rows of object detection predictions across synthetic driving scenes.
Detection inference examples published in the team repository. Source ↗

The problem

Training perception models requires images and reliable annotations. Synthetic scenes make it possible to control camera configurations and produce multiple ground-truth representations from the same scene.

My contribution

Co-engineered the rendering pipeline and built hooks to capture raw frame buffers. Automated annotation export into COCO JSON and YOLO text formats, and authored GLSL fragment shaders for synchronous RGB, depth, and instance segmentation outputs.

Evaluation

Benchmarked YOLOv8s, RT-DETR-L, and Faster R-CNN to examine cross-dataset generalization and the trade-off between detection accuracy and inference latency. Observed results span 0.694–0.934 mAP@50 and 8.2–22.0 ms across evaluated configurations; these are not a single-model before/after improvement.

The important limitation

Synthetic data is not a substitute for real-world evaluation. Rendering assumptions, scene diversity, and camera geometry all affect transfer. Results need to be interpreted within the evaluated dataset and hardware setup.

Next project

BKEL Updater