DispNet-TensorFlow

TensorFlow implementation of A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation by Zhijian Jiang.

Dataset

• Scene Flow Datasets: FlyingThings3D, Driving, Monkaa

Tutorials

TensorFlow

• Tensorflow tutorials (Eng Sub) 神经网络 教学 教程

Network

Convolutional Network

Name	Kernel	Strides	Channels I/O	Input Resolution	Output Resolution	Input
conv1	7 * 7	1	6/64	1536 * 768	768 * 384	Images
max_pool1	2 * 2	2	64/64	1536 * 768	768 * 384	conv1
conv2	5 * 5	1	64/128	768 * 384	384 * 192	max_pool1
max_pool2	2 * 2	2	128/128	768 * 384	384 * 192	conv2
conv3a	5 * 5	1	128/256	384 * 192	192 * 96	max_pool2
max_pool3	2 * 2	2	256/256	384 * 192	192 * 96	conv3a
conv3b	3 * 3	1	256/256	192 * 96	192 * 96	max_pool3
conv4a	5 * 5	1	256/512	192 * 96	96 * 48	conv3b
max_pool4	2 * 2	2	512/512	96 * 48	96 * 48	conv4a
conv4b	3 * 3	1	512/512	96 * 48	96 * 48	max_pool4
conv5a	5 * 5	1	512/512	96 * 48	48 * 24	conv4b
max_pool5	2 * 2	2	512/512	48 * 24	48 * 24	conv5a
conv5b	3 * 3	1	512/512	48 * 24	48 * 24	max_pool5
conv6a	5 * 5	1	512/512	48 * 24	24 * 12	conv5b
max_pool6	2 * 2	2	1024/1024	24 * 12	24 * 12	conv6a
conv6b	3 * 3	1	1024/1024	24 * 12	24 * 12	max_pool6
pr6 + loss6	3 * 3	1	1024/1	24 * 12	24 * 12	conv6b

Upconvolutional Network

Name	Kernel	Strides	Channels I/O	Input Resolution	Output Resolution	Input
upconv5	4 * 4	2	1024/512	24 * 12	48 * 24	conv6b
iconv5	3 * 3	1	1024/512	48 * 24	48 * 24	upconv5 + conv5b
pr5+loss5	3 * 3	1	512/1	48 * 24	48 * 24	iconv5
upconv4	4 * 4	2	512/256	48 * 24	96 * 48	iconv5
iconv4	3 * 3	1	768/256	96 * 48	96 * 48	upconv4 + conv4b
pr4+loss4	3 * 3	1	512/1	96 * 48	96 * 48	iconv4
upconv3	4 * 4	2	256/128	96 * 48	192 * 96	iconv4
iconv3	3 * 3	1	384/128	192 * 96	192 * 96	upconv3 + conv3b
pr3+loss3	3 * 3	1	128/1	192 * 96	192 * 96	iconv3
upconv2	4 * 4	2	128/64	192 * 96	384 * 192	iconv3
iconv2	3 * 3	1	192/64	384 * 192	384 * 192	upconv2 + conv2
pr2+loss2	3 * 3	1	64/1	384 * 192	384 * 192	iconv2
upconv1	4 * 4	2	64/32	384 * 192	768 * 384	iconv2
iconv1	3 * 3	1	96/32	768 * 384	768 * 384	upconv1 + conv1
pr1+loss1	3 * 3	1	32/1	768 * 384	768 * 384	iconv1

Issues

• How to input png images:

  • Solution 1 -- FAIL:

   contents = ''
   with open('path/to/image.jpeg') as f:
  		contents = f.read()
   tf.image.decode_jpeg(contents) 
  

  • Solution 2 -- FAIL

   reader = tf.WholeFileReader(http://stackoverflow.com/questions/34340489/tensorflow-read-images-with-labels)
   key, value = reader.read(filename_queue)
   example = tf.image.decode_png(value)
  

  • Solution 3 -- Success

   file_contents = tf.read_file(input_queue[0])
   example = tf.image.decode_png(file_contents, channels=3)