I aim to train a pedesterian detector. However, as far as I see from the examples, I need to keep the aspect ration of the bounding boxes exactly with the detection window size to train a detector. However, this is not a easy task. There are many different shapes and sizes of bounding boxes. What is the handy way to deal with that constraint?
Best
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Here's my python script that takes XML output of tools/imglab and modifies the height of boxes so that they all are of equel aspect ratio: http://pastebin.com/3ecMShu2
Now, the same functionality is probably somewhere in dlib but I didn't find it when using imglab and train_object_detector sample.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I aim to train a pedesterian detector. However, as far as I see from the examples, I need to keep the aspect ration of the bounding boxes exactly with the detection window size to train a detector. However, this is not a easy task. There are many different shapes and sizes of bounding boxes. What is the handy way to deal with that constraint?
Best
You need to train different detectors for different aspect ratios. I would
pick maybe 4 or 5 different sizes and go with those.
I also faced the same problem.
Here's my python script that takes XML output of tools/imglab and modifies the height of boxes so that they all are of equel aspect ratio: http://pastebin.com/3ecMShu2
Now, the same functionality is probably somewhere in dlib but I didn't find it when using imglab and train_object_detector sample.