This repository contain source code to create the dataset for the content-aware DCNN.
NOTE:
- download the weights for SSD model here or here and put into the ./weights directory
- download the imagenet_labels and put into root directory. This file contains all annotation data.
- download the youtube-bb dataset
- clone the repository to download and segment youtube videos.
General steps STEP1: process raw youtube vid from the file yt_bb_detection_train.csv, filter out only videos with 1080x1920p resolutions. Then use the source code at (4) to download and segment the videos. NOTE: the source code to download & segment youtube-bb dataset has a bug, when the youtube-dl only download 720p videos at max resolution To fix this issuse, remove best parameter, and adding -c to the end of the script. Location: inside function dl_and_cut, located in the youtube_bb.py file
STEP2: split the video segment The videos are localed inside ./data/YOUTUBE_data/videos, store in separated directory. Each directory is named after the class ID. Currently, there are 13 class IDs: 0: PERSON, 1: BIRD, 2: BICYCLE, 3: WATERCRAFT, 4: BUS, 6: COW, 7: CAT, 10: HORSE, 11: MOTORCYCLE, 13: AIRPLANE, 15: TRAIN, 19: DOG, 23: CAR run ./scripts/split_vid.sh to split the videos into multiple bitrates and resolutions.
STEP3: split transformed videos into image frames