Back to catalogue
Project
COCO Object Detection Demo
OpenCV integration of a pretrained COCO detector with image preprocessing and inference.
How it’s builtHow it’s built
COCO Object Detection Demo · OpenCV DNN webcam demo of a COCO-trained SSD MobileNet v2 detector
I wrote OpenCV scripts that load a pretrained SSD MobileNet v2 COCO model through cv2.dnn, run it on webcam frames or still images at 300x300, and draw labelled boxes above a 0.25 confidence threshold. A second script runs OpenCV's ResNet-10 SSD face detector on a video stream. The pretrained weights come from TensorFlow and OpenCV, not my own training.
- lines of own Python across 4 scripts
- ~370
- detector input size
- 300x300
- confidence thresholds (COCO / face)
- 0.25 / 0.7
Technology and what it does here
- OpenCV DNN
- ReadNetFromTensorflow loads frozen_inference_graph.pb with the SSD MobileNet v2 pbtxt
- SSD MobileNet v2 (COCO)
- Pretrained detector, labels from coco-labels-2014_2017
- OpenCV ResNet-10 SSD (Caffe)
- Face detection in program.py from res10_300x300_ssd fp16 weights
- NumPy / matplotlib
- Image channel exploration and still-image result display
Implementation details
- app.py runs live webcam detection and draws filled label backgrounds with class and score for each box above threshold.
- tf_text_graph_ssd.py is OpenCV's own script for generating the .pbtxt graph and should not be counted as my code.
Technologies
- OpenCV
- COCO
- Object Detection
