jiahaopang / crl Goto Github PK

I test the pretrained model on SceneceFlow and Middlebury2014 datasets， but the result is not good. Maybe the pretrained model is only for the Kitti2015 datasets. Could you explain it? @Artifineuro
Or could you please release your pretrained model on scene flow and middlebury dataset?

你这个coord_canon里面的数值应该是记录着点的坐标吧。比如坐标(1,1)的点，它的值就是(1,1)？

Is there a python version for inference. I am able to run a forward pass with the following code. Output is shown.

For inference, is net.blobs('zero').set_data(zeros(width_test, height_test));
net.blobs('coord_canon').set_data(coord); needed?

path_left = 'left.jpg'
path_right = 'right.jpg'
MODEL_FILE = 'deploy_kitti.prototxt'
PRETRAINED = 'crl.caffemodel'
left_img = cv2.imread(path_left)
right_img = cv2.imread(path_right)
caffe.set_mode_gpu()
caffe.set_device(0)
net = caffe.Net(MODEL_FILE, PRETRAINED, caffe.TEST)
left_img = cv2.resize(left_img, (net.blobs['img0'].data.shape[3],net.blobs['img0'].data.shape[2]))
right_img = cv2.resize(right_img, (net.blobs['img1'].data.shape[3],net.blobs['img1'].data.shape[2]))
left_img = left_img.astype(np.float32)
left_img = left_img.transpose((2, 0, 1))
right_img = right_img.astype(np.float32)
right_img = right_img.transpose((2, 0, 1))
net.blobs['img0'].data[...] = left_img
net.blobs['img1'].data[...] = right_img
caffe.set_mode_gpu()
caffe.set_device(0)
disp = net.forward()
out = -disp["predict_flow2_s2"].transpose(2,3,0,1).reshape(320,1792)
img = Image.fromarray(out, 'RGB')
img.save('my.png')
img.show()

I believe some cmake files are missing?

I'm trying to reproduce the results of CRL, but when I run the test_kitti.m on matlab, I'm having the error bellow:

[...]
predict_flow3_s2_0_split
I0322 15:41:00.695804 2244 net.cpp:753] Ignoring source layer Downsample4_s2
I0322 15:41:00.695807 2244 net.cpp:753] Ignoring source layer flow_loss3_s2
I0322 15:41:00.695911 2244 net.cpp:753] Ignoring source layer Convolution8_s2_Convolution8_s2_0_split
I0322 15:41:00.695919 2244 net.cpp:753] Ignoring source layer Downsample5_s2
I0322 15:41:00.695920 2244 net.cpp:753] Ignoring source layer flow_loss2_s2
W0322 15:41:00.894263 2244 net.hpp:41] DEPRECATED: ForwardPrefilled() will be removed in a future version. Use Forward().
F0322 15:41:03.907975 2244 syncedmem.cpp:64] Check failed: error == cudaSuccess (2 vs. 0) out of memory
*** Check failure stack trace: ***
Killed

This error appears on the prompt, since I opened matlab with the prompt, and the error happens in the line of the test_kitti.m file on matlab:

net.forward_prefilled();

It crashes the program and force me to close Matlab.
Has anyone ever had this error and knows a solution?
Tks.

Hello,
I am trying to re-implement this network in TensorFlow, but I haven’t reached this precision. And some details，especially data augmentation process ， of training process is a bit confusing.
Could you possibly list the detail of data augmentation，and if you make a horizontal scale transformation in training process.

Hello,
I am currently trying to re-implement your network in PyTorch. Could you possibly list what kind of data augmentation (with parameters) you use, so that I could reproduce it. You did not mention it in the paper so it is a bit confusing.. Thank you!

when I set single_direction to be -1, once run, a cuda error will be output. Anyone knows the reason?

我在 这个网站 http://ethereon.github.io/netscope/#/editor 上对网络模型进行可视化，会出现以下错误

Error Encountered
Uncaught TypeError: Cannot set property 'class' of undefined
Warning
Can't infer network data shapes. Can't infer output shape of the 'img0s_aug' layer of type 'DataAugmentation'. TypeError: Cannot read property 'batch_size' of undefined

thanks for the codes. the idea that using ResNet to rectify the results is wonderful!
i have two problems.

1. in 'correlation 1d layer', have you try to set the "single_direction" -1 ? due to the samples all have the positive disparity. i do the test and find the result does not achieve expectation.
2. have you try to generate the 'cost_volume' using caffe? i found it is very hard using caffe. if you do a success trial . Please let me know! Thx~