Computer Vision Onramp - Part 4 - Training and Postprocessing

Detect and Count Objects

Choose a Detection Algorithm

Object detection training algorithms fall into two categories:

1. Machine Learning (e.g., ACF): Faster, require less data, detect one class at a time. Good for consistent viewpoints.
2. Deep Learning (e.g., YOLO): More accurate, handle more variability and multiple classes, but require more data and a GPU.

ACF (Aggregated Channel Features) is ideal for turtles because they don’t change significantly in appearance.

Train an ACF Detector

detector = trainACFObjectDetector(imsWithBoxLabels, NumStages=5);
[bbox, score] = detect(detector, frame);

If the detector mislabels objects (like a rock), increase the NumStages (e.g., to 20) to improve accuracy.

Postprocess Detections

1. Thresholding Scores

Remove low-confidence detections. Use a histogram to determine a threshold:

histogram(score, 10)
bbox = bbox(score > 30, :);
score = score(score > 30);

2. Non-Maximum Suppression (NMS)

Remove overlapping boxes identifying the same object using selectStrongestBbox:

[selectedBbox, selectedScore] = selectStrongestBbox(bbox, score, OverlapThreshold=0.1);

Count and Annotate

numBoxes = size(selectedBbox, 1);
str = numBoxes + " turtle(s) detected";
imgCounted = insertText(frame, [250 550], str);
imshow(imgCounted)

Apply to Each Video Frame

Use a while loop with hasFrame to process a video stream:

while hasFrame(turtleVideo)
    frame = readFrame(turtleVideo);
    % ... detect, postprocess, and count ...
    imshow(img)
    drawnow
end