I use a third party computer vision library to detect objects from stream of images. That library use float to store dimensions of bounding boxes and it doesn't provide bounding boxes dimensions in higher precision (e.g., double). I'm interested in the diagonals of bounding boxes enclosing detected objects. I used the hypot() function in <math.h> to calculate the diagonal of bounding boxes from their heights and widths. I save my outputs (height, width, diagonal of bounding boxes) to json files. I plot histograms of dimensions of bounding boxes (read from saved json files), but I keep getting a weird trough in the resulting histograms of diagonals at a specific bin: Sample 1: The trough is the bin range between 1.19 and 1.26 in Figure 3. Sample 2: The trough is the bin range between 1.19 and 1.26 in Figure 8. I suspect the troughs is caused by how I calculate diagonal from height and width because histograms of height and width doesn't have those troughs. So I plot the height and width of bounding boxes in those troughs to check if there is any abnormality (infinity, subnormal) but there isn't any: Sample 1: Sample 2: I appreciate any pointer to how I should go about investigating those weird troughs. Continue reading...
