DepthEye12 Depth-Aware Object Detection¶
Completed
Overview¶
DepthEye12 is a computer vision project I developed to combine object detection with monocular depth estimation.
The idea was to associate each detected object with an estimated depth value using a single RGB camera, without requiring a dedicated depth sensor.
Implementation¶
I combined two pretrained models:
- YOLOv12 for object detection and bounding box prediction.
- MiDaS DPT-Hybrid for monocular depth estimation.
For each video frame, YOLOv12 identifies objects while MiDaS generates a depth map.
I then sample the depth values within each detected bounding box to associate a depth estimate with the corresponding object.
Both models are loaded into GPU memory once and reused throughout video processing.
The pipeline is implemented in Python using PyTorch and OpenCV, with a DetectionDepthModel class handling both models.
Results and Limitations¶
The completed pipeline processes video files and produces annotated videos containing bounding boxes, class labels, confidence scores, and estimated depth values.
However, MiDaS provides relative depth estimates rather than calibrated physical distances. The resulting values cannot be directly interpreted as meters without additional calibration.
I did not record a quantitative evaluation of depth accuracy or inference throughput, so the project does not establish metric distance accuracy or real-time performance.