Skip to content
~/projects/rcnn
All projects
shippedComputer Vision2024

R-CNN Object Detection Studio

R-CNN object detection with bounding-box visualization and an interactive Streamlit interface for inspecting predictions, applied to marine-debris detection.

Source
problem

The problem

A detection model is only trustworthy if you can *see* what it sees. Raw mAP numbers hide the failure cases; you need to inspect predictions frame by frame, especially for something like marine debris where classes are subtle.

approach

How I approached it

  • Built R-CNN-based object detection with bounding-box visualization in PyTorch + OpenCV.
  • Deployed an interactive Streamlit interface to visually inspect predictions and datasets, an artist/analyst-facing tool.
  • Focused on tight feedback loops between AI output and human review.
architecture

Architecture

1Propose

Region proposals over the input image

2Classify

CNN scores each region + refines boxes

3Visualize

Draw boxes + labels over the frame

4Inspect

Streamlit UI to step through predictions

decisions

Engineering decisions & tradeoffs

$

Streamlit inspection UI

// Making predictions visually inspectable turned a black-box model into something a human could audit and trust.

$

Marine debris as the target

// A real, hard, socially useful detection problem, small, cluttered, low-contrast objects.

stack
PythonPyTorchOpenCVStreamlit
what's next
  • › Swap R-CNN for a single-stage detector (YOLO-class) for live video.
  • › Active-learning loop: send low-confidence frames back for labeling.
All projectsDiscuss this work