search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.75k stars 7.96k forks source link

Interpreting the YOLO output #7368

Open oo92 opened 3 years ago

oo92 commented 3 years ago

I am currently training on a dataset and I want to make sense of the output from YOLOv4-tiny.

v3 (giou loss, Normalizer: (iou: 0.50, obj: 1.00, cls: 1.00) Region 30 Avg (IOU: 0.000000), count: 1, class_loss = 0.000007, iou_loss = 0.000000, total_loss = 0.000007 v3 (giou loss, Normalizer: (iou: 0.50, obj: 1.00, cls: 1.00) Region 37 Avg (IOU: 0.773155), count: 5, class_loss = 1.460724, iou_loss = 3136.319092, total_loss = 3137.779785 total_bbox = 38508, rewritten_bbox = 0.000000 % v3 (giou loss, Normalizer: (iou: 0.50, obj: 1.00, cls: 1.00) Region 30 Avg (IOU: 0.000000), count: 1, class_loss = 0.000012, iou_loss = 0.000000, total_loss = 0.000012 v3 (giou loss, Normalizer: (iou: 0.50, obj: 1.00, cls: 1.00) Region 37 Avg (IOU: 0.800353), count: 3, class_loss = 1.261792, iou_loss = 1423.418457, total_loss = 1424.680298 total_bbox = 38511, rewritten_bbox = 0.000000 %

  1. What is iou:0.50? It seems to me like a threshold of some kind? Why is it there? As well as obj and cis?

  2. What is count? Is it counting the classes in an image?

  3. Why is my top IOU 0.00000 but the bottom IOU is 0.773? Why is the top just 0s but the bottom one is higher? Its started to do this only when I began using a new dataset.

  4. What is Region 30, 37?

  5. What would make rewritten_bbox 0.00000%? Why is total bbox so high when I only have, at most, 3 objects in an image?

asfarley commented 3 years ago

iou is intersection over union, it's a bounding-box error metric. Not sure about the rest.