tonnyesp/deep-stereo

Use deep learning to estimate virtual views

Useful commands

Play ballet depht camera

mplayer -demuxer rawvideo -rawvideo w=1024:h=768:format=y8 ballet-depth_2.yuv

Play ballet color camera

mplayer -demuxer rawvideo -rawvideo w=1024:h=768:format=i420 ballet-color_2.yuv

Generate solid green image

convert -size 1024x768 canvas:green canvas_green.png

Generate solid black image

convert -size 1024x768 xc:#FFFFFF red.png

Generate green canvas video to display

ffmpeg -loop 1 -i canvas_green.png -c:v libx264 -t 30 -pix_fmt yuv420p out.mp4

Generate a 100 frame video using image on canvas_green.png

ffmpeg -loop 1 -i canvas_green.png -c:v rawvideo -r 1 -t 100 -pix_fmt yuv420p out.yuv

Note: Change -pix_fmt to y8 to convert depth estimation videos

Use 1 gpus in caffe

--gres=gpu:1

WHICH CNN Frameworks to use:

* Theanos
* Keras (Prefered, new)
* Tensorflow (from Google)

Extract 1 frame from undo dancer seq

ffmpeg -ss 00:00:05 -t 1 -s 1920x1088 -i Dancer_c_2_1920x1088.yuv -f mjpeg Dancer_c2_frame.jpg

Camera merge test sequences for plane sweep volume

./merge_cameras.sh ../../sweep_results_test/cam0 ../../sweep_results_test/cam1 ../../sweep_results_test/merged

Search for variables defined in TF variable scope

[n.op.name for n in tf.all_variables()]