suyogduttjain / fusionseg Goto Github PK

I've emailed you but I didn't receive any reply. So I have to raise an issue here. Hope it won't bother you too much.
I want to use your YOUTUBE OBJECTS DATASET (PIXEL LEVEL OBJECT MASKS). But I cannot match it with the original youtube objects dataset.
For example, how can I find the image "ROOT OF masks\youtube_masks\aeroplane\data\0012\shots\001\images\00001.jpg" in the original youtube objects dataset? I tried to match this image with the first image in video "0012" ,shot "001", using the range files from the original dataset. But it's wrong.
I know it's a stupid question but I really cannot start my experiment without the matching.
Thank you!

Can you share your prototxt file for training appearance and motion stream, I am facing issues in loss layer for training appearance stream.

Im using caffe with gpu from docker (caffe version 1.0.0)
And I get the follwing error:

root@f388bca8865f:/fusionseg# python fusionseg.py appearance jpg ./sample/
WARNING: Logging before InitGoogleLogging() is written to STDERR
W0302 12:35:03.005851 622 _caffe.cpp:139] DEPRECATION WARNING - deprecated use of Python interface
W0302 12:35:03.005897 622 _caffe.cpp:140] Use this instead (with the named "weights" parameter):
W0302 12:35:03.005903 622 _caffe.cpp:142] Net('/fusionseg/models/appearance/deploy.prototxt', 1, weights='/fusionseg/models/appearance/appearance_stream.caffemodel')
[libprotobuf ERROR google/protobuf/text_format.cc:274] Error parsing text-format caffe.NetParameter: 7244:14: Message type "caffe.LayerParameter" has no field named "interp_param".
F0302 12:35:03.010118 622 upgrade_proto.cpp:88] Check failed: ReadProtoFromTextFile(param_file, param) Failed to parse NetParameter file: /fusionseg/models/appearance/deploy.prototxt
*** Check failure stack trace: ***
Aborted (core dumped)

Hi,

I tried to adapt the appearance stream of your code on other datasets but wasn't able to make it trainable. I tried training with SGD and Adadelta, neither worked. Also the weights that I tried was ResNet101 ImageNet pretrained model and your pretrained appearance model.
I've attached my solver and training prototxt here, I'd appreciate if someone could help me figure out what went wrong.
https://gist.github.com/vincentzhang/2f6f5bbb81313470e023c43c246ea5b0

Thanks

...

Recently I have been studying video object segmentation got a chance to read your paper “FusionSeg: Learning to combine motion and appearance for fully automatic segmentation of generic objects in videos”. After reading it carefully, I still have some problem that how to preprocess the dataset and train this caffemodel. I would appreciate it if you could release the training dataset and the source code to process the data proposed in the paper.
I want to implement the process of model training. Thanks for your generosity.

I want to evaluate the joint model.
Thanks.

Hello,

I was trying to run python demo.py motion but I got this error:
caffe.bin: not found
Note I have modified demo.py to point to the caffe.bin in deeplab-public-ver2/build/tools/caffe.bin but it still gave the same error.

Thanks.

Once you have segmented the objects what can you do with that information? Is there an output I can use?

how to obtain "Optical flow encode as RGB images"?