all projects
RGB-D

3D Object Detection on RGB-D

An end-to-end 3D object-detection pipeline on RGB-D data, comparing a from-scratch model against a fine-tuned EfficientNet-B3 backbone under a small-data regime.

Result 8-corner boxes → parametric (center + size + quaternion) via PCA Stars ★ 1 Tags RGB-D · 3D bbox · EfficientNet · Multimodal
View on GitHub Pipeline visualization

A compact, multimodal 3D detection pipeline that learns oriented bounding boxes from RGB-D input with only ~200 training samples — a deliberate small-data challenge.

Approach

  • Two backbones compared. A from-scratch network vs a fine-tuned EfficientNet-B3.
  • Box parameterization. Eight corner points are converted to a clean parametric form — center, size and orientation (quaternion) — using PCA, which stabilizes the regression target.
  • Interactive pipeline view. The repo ships an HTML visualization of each stage, from RGB-D input to predicted boxes.