bryanyzhu / deepof Goto Github PK

Thank you for your great job!
I have two questions:

1. You said your tensorflow implementation are lower than pytorch and caffe, does it matter a lot?
   2.In your paper "Hidden Two-Stream Convolutional Networks for Action Recognition" , you use a much smaller architecture than flownetS, do you test a smaller architecture in this paper? and the corresponding EPE ,fps ?
   thank you

Hi, I am working on estimating optical flow using cnn and I found your excellent work. I have opened several issues in gluon-cv, so probably u remember me.

My issue is the warp operation of image. Basically, warp means input image1 and optcial flow, output image2. And u said in the paper that

The inverse warp is performed using a spatial transformer module [13] inside the CNN.

I try to find the following code, which seems not related to spatial transformer:

so, probably I didn't find the right code, and could u tell me where's the inverse warp operation in your code? Also, I am interesting about the above code, can u give some references about image warp as I counldn't fully understand some lines in the warp operation.
Thanks in advance!

Hello,
I want to replicate it for my own dataset. The codes listed here are very difficult to understand. Could you please include a read me file.

Dear Dr.Zhu
I am interested in deepOF and I downloaded your codes. However, I found that the step "assert(os.path.exists(img_path))" reports an error. It seems that this means that there are no img_paths, which is used for storing the training samples. May I know where to download there samples?
Thanks in advance.

Dear Dr.Zhu
I have read through your paper and codes using Sintel for training last week. In your paper, FlowFields is used as the classical optical flow estimator to provide the proxy ground truth. This proxy groundtruth is used to train the CNN-based optical flow estimators. According to my understanding, it seems that the proxy groundtruth are the pr1~6 in the codes, for they are used as the "flows" in the function "loss_interp_multi(flows, inputs, epsilon, alpha_c, alpha_s, lambda_smooth, flow_scale, deltaWeights)", as shown in the codes below:

(in function "inception_v3" in sintelWrapFlow.py): 
loss6, _ = loss_interp_multi(pr6, pr6_input, epsilon, alpha_c, alpha_s, lambda_smooth, flow_scale_6, deltaWeights)

(in function "loss_interp_multi" in sintelWrapFlow.py):
def loss_interp_multi(flows, inputs, epsilon, alpha_c, alpha_s, lambda_smooth, flow_scale, deltaWeights):
...
flows = tf.multiply(flows, flow_scale)
flows_flat = tf.reshape(flows, [num_batch, -1, 2])
floor_flows = tf.to_int32(tf.floor(flows_flat))
weights_flows = flows_flat - tf.floor(flows_flat)

However, it seem that the pr6 is computed from end_points["Mixed_7c"], as shown in the codes below:

(in function "inception_v3" in sintelWrapFlow.py): 
pr6 = slim.conv2d(end_points["Mixed_7c"], 2*(time_step-1), [3, 3], activation_fn=None, scope='pr6')

However, end_points["Mixed_7c"] is obtained using function "inception_v3_base". It seems that function "inception_v3_base" is not an "non-CNN based classical methods", because convolution and filtering layers are used in this function. Could you help me find where did I mis-understand your code? Thank you very much!