bjou

Deep Cross Residual Learning for Multitask Visual Recognition

Authors: Brendan Jou and Shih-Fu Chang
Institution: Digital Video & Multimedia Lab @ Columbia University

Last verified compatible Caffe version: f28f5ae2f2453f42b5824723efc326a04dd16d85

Best published X-ResNet model on the Multitask-structured VSO dataset (from Table 2 in our paper):

Model

bjou

#!/bin/sh

# check arg inputs
verbose=false
while getopts "hi:o:s:d:p:v" opt; do
    case $opt in
        h)
            echo ""
            exit 1
            ;;

bjou

# convert HD video to SD and transcode in one-shot
ffmpeg -i video.ts -preset ultrafast -y [-loglevel quiet] -aspect 16:9 -s 640:360 video.mp4

# batch segment the video in one-shot (no looped seeking)
# NOTES: "-vcodec" and "-acodec" can be changed to preferred stream types, e.g. libx264 and libfaac, respectively
#        "-vbsf h264_mp4toannexb" optionally allows for broken mp4's to still get parsed
ffmpeg -i video.mp4 -vcodec copy -acodec copy -map 0 [-vbsf h264_mp4toannexb] [-loglevel quiet] -f segment -segment_format mp4 -segment_times <time_sec_1>,<time_sec_2>,<time_sec_3>,<...> output%02d.mp4

# TODO: needs further investigation on segmentation accuracy; cutting may still happen on i-frames due to keyframe-level accuracy, but this may have been solved by mpegts-to-mp4 transcoding

bjou

for i in *.mp4; do ffmpeg -y -i "$i" -aspect 16:9 -s 640:360 "${i%.*}"_sd.mp4; done

bjou

ffmpeg -i input_file.mp4 -acodec copy -vcodec copy -f mov output_file.mov

bjou

ffmpeg -i "$inFile" -vn -acodec pcm_s16le -ar 16000 -ac 1 -f s16le -y "$rawFile"

bjou

ffmpeg -i INPUT -ss TIME -t DURATION -preset ultrafast -y [-loglevel quiet] OUTPUT.mp4

bjou

ffmpeg -framerate 3 -i %05d.png -c:v libx264 -r 3 <MP4FILE>.mp4

ffmpeg -i <MP4FILE>.mp4 -c:v libx264 -preset fast \
       -pix_fmt yuv420p -r 23.98 -g 15 -s <W>x<H> -b 2500k -bt 300k \
       -f mov <OUTFILE>.mov