shenjianbing / pdb-convlstm Goto Github PK

Don't you need to add some extra convs to help output saliency during pre-training spatial-learning part?

Hi, @shenjianbing , I am trying to run your code. However, I could not fine DAVIS 2016 test set anywhere, the website only presents the train and val set. Can you share the test set?

You said the DB-ConvLSTM using a 32 channel filter with 1,2 dilations, kernel size of 3x3.
And you must concatenate 2(fw, bw) outputs from each branch -> CAT [(B, T, 32, 60, 60),(B, T, 32, 60, 60), 2] = (B, T, 64, 60, 60). But you did not mention, how does it use another Conv Layer to filter the concatenated result. ie. Kernel sizes, dilations.

This work looks pretty good and impresses me a lot. However, I am confused about the loss functions. Is there any reason why you use the combination of cross-entropy loss and MAE loss? There is not enough explanation in the paper.

Besides, if you train the proposed method only with cross-entropy loss or MAE loss individually, what is the performance?

Hello,

And thanks for sharing your work. I'm trying to run your network, replaced paths in davis_path.txt, seq_t5_n1.txt. I also tried replacing paths in test.prototxt. Unfortunately, there is a weird error about an unknown path. Could you help me? Thanks!

Load net...
WARNING: Logging before InitGoogleLogging() is written to STDERR
W0122 16:14:50.405848 90842 _caffe.cpp:135] DEPRECATION WARNING - deprecated use of Python interface
W0122 16:14:50.405881 90842 _caffe.cpp:136] Use this instead (with the named "weights" parameter):
W0122 16:14:50.405889 90842 _caffe.cpp:138] Net('/home/user/PDB/test.prototxt', 1, weights='/home/user/PDB/model/pdb-convlstm.caffemodel')
I0122 16:14:50.408905 90842 net.cpp:51] Initializing net from parameters:
name: "May_lstm"
state {
phase: TEST
level: 0
}

[...]

layer {
name: "conv4_6_1x1_reduce"
type: "Convolution"
bottom: "conv
I0122 16:19:14.291399 91324 layer_factory.hpp:77] Creating layer data
I0122 16:19:14.291419 91324 net.cpp:84] Creating Layer data
I0122 16:19:14.291422 91324 net.cpp:380] data -> data
I0122 16:19:14.291438 91324 net.cpp:380] data -> fakelabel
I0122 16:19:14.291450 91324 image_data_layer.cpp:38] Opening file /home/user/PDB/my_path.txt
I0122 16:19:14.291590 91324 image_data_layer.cpp:63] A total of 588 images.
I0122 16:19:14.373279 91324 image_data_layer.cpp:90] output data size: 5,3,473,473
I0122 16:19:14.396646 91324 net.cpp:122] Setting up data
I0122 16:19:14.396672 91324 net.cpp:129] Top shape: 5 3 473 473 (3355935)
I0122 16:19:14.396682 91324 net.cpp:129] Top shape: 5 (5)
I0122 16:19:14.396685 91324 net.cpp:137] Memory required for data: 13423760
I0122 16:19:14.396690 91324 layer_factory.hpp:77] Creating layer clip_markers
I0122 16:19:14.396708 91324 net.cpp:84] Creating Layer clip_markers
I0122 16:19:14.396713 91324 net.cpp:380] clip_markers -> reshape-cm
I0122 16:19:14.396723 91324 hdf5_data_layer.cpp:80] Loading list of HDF5 filenames from: /home/user/PDB/maycaffe-convlstm/models/seq_t5_n1.h5
I0122 16:19:14.396755 91324 hdf5_data_layer.cpp:94] Number of HDF5 files: 6
HDF5-DIAG: Error detected in HDF5 (1.10.0-patch1) thread 139706439980672:
#000: ../../../src/H5F.c line 579 in H5Fopen(): unable to open file
major: File accessibilty
minor: Unable to open file
#1: ../../../src/H5Fint.c line 1100 in H5F_open(): unable to open file: time = Wed Jan 22 16:19:14 2020
, name = '�HDF', tent_flags = 0
major: File accessibilty
minor: Unable to open file
#2: ../../../src/H5FD.c line 812 in H5FD_open(): open failed
major: Virtual File Layer
minor: Unable to initialize object
#3: ../../../src/H5FDsec2.c line 348 in H5FD_sec2_open(): unable to open file: name = '�HDF', errno = 2, error message = 'No such file or directory', flags = 0, o_flags = 0
major: File accessibilty
minor: Unable to open file
F0122 16:19:14.397401 91324 hdf5_data_layer.cpp:31] Failed opening HDF5 file: �HDF
*** Check failure stack trace: ***
Abandon (core dumped)

Hi,
I have tried to compile your caffe version on both cuda 8.0, cudnn 7.4 and cuda 9.2, cudnn 7.4. My compilation is successful but when I run test_davis.py I get this error:
F1121 00:18:51.439702 30494 common.cpp:152] Check failed: error == cudaSuccess (35 vs. 0) CUDA driver version is insufficient for CUDA runtime version

I have tried this on multiple machines but error remains the same. Can you please guide me which version do you use?

Recently, I am studying your paper "Local Semantic Siamese networks for fast tracking". The paper says that "our source code is available at https://github.com/shenjianbing/LSSiam", but I didn't found this website. Would you provide a right web to me? Thank you very much!!!

typo...

您好请问您，这篇文章模型是如何训练的，在您公开的代码里没有找到相关的readme，您可不可以说明下呢？谢谢

Hello, your project is very attractive to me, because I want to test smoke saliency, so I want to train on my own dataset, can you provide training code? Thank you very much.

Thank you very much for sharing your excellent work, and I would like to ask two questions about the FBMS dataset. The first is how do you deal with multiple object sequences in the test set, and do you treat them all as foreground segmentation? Second, did you use the Official DAVIS evaluation code to evaluate the FBMS dataset?

🔥
You used the GT labels from the training set, why you could claim yourself unsupervised segmented the primary object?
BTW, maybe I have some mis-understanding about the DAVIS unsupervised TEASER challenge.
Did it mean when inferencing, the given data is only video sequence. And we could use the GT labels as training label? (I think it is not the so-called unsupervised learning.)

Thank you for your interesting work, which line of your code embodies parallel dilated convo�lution branches?

Hi. Can you provide the predicted maps of Youtube-Object?