convert the format of the caltech pedestrian dataset to the format that yolo uses

This repo is adapted from

• https://github.com/mitmul/caltech-pedestrian-dataset-converter
• https://pjreddie.com/media/files/voc_label.py

dependencies

• opencv
• numpy
• scipy

how to

1. Convert the .tar files with video to .png frames by running $ ./generate-images.py {path_to_derectory_with_tar_files} {path_to_output} [amount_of_threads]. This script takes all files from <path_to_directory_with_tar_files>, extracts each archive and converts extracted .seq files to <path_to output>, finally it removes folders with extracted .seq files. Parallel mode provides way to go through above described flow in regime like "one thread per archive". To enable parallelism need just to specify <amount_of_threads> as integer value. The most efficient way is to specify number of threads equal to number of the archives in directory (sure if you have enough cores on CPU).
2. Squared images work better, which is why you can convert the 640x480 frames to square frames by running $ python squarify-images.py {path_to_derectory_with_images} {output_directory_path} {size_of_square_side} [amount_of_threads]. It converts images to 640x640(by adding white bar) and then to specified size.
3. Convert the .vbb annotation files to .txt files by running $ python generate-annotation.py {path_to_annotation.zip} {train_samples_output_directory} {test_samples_output_directory} [squarified_image_side_length]. It will create test.txt and train.txt into each output directory respectively. Also it will create bunch of .txt files named as images connected with them. For images wchich don't contain any lables .txt files won't be created.
4. Adjust .data yolo file
5. Adjust .cfg yolo file: take e.g. yolo-voc.2.0.cfg and set height = {height_of_your_img}, width = {width_of_your_img}, classes = 2, and in the final layer filters = 21 (= (classes + 5) * 3))

folder structure

|- caltech
|-- annotations
|-- test06
|--- V000.seq
|--- ...
|-- ...
|-- train00
|-- ...
|- caltech-for-yolo (this repo, cd)
|-- generate-images.py
|-- generate-annotation.py
|-- images
|-- labels
|-- test.txt
|-- train.txt