search

jutanke / cabbage

Unofficial Implementation of the paper: Multiple People Tracking by Lifted Multicut and Person Re-identification
MIT License
93 stars 26 forks source link

original paper using deepcut to fuse body part info #7

Open shensheng27 opened 6 years ago

shensheng27 commented 6 years ago

From the code i cannot find body part fusing. Is the fusing step in provided h5 model?

chenyuan1994 commented 6 years ago

I think fusing body part information is the property of the StackNet @shensheng27

shensheng27 commented 6 years ago

@chenyuan1994 Thanks for reply. I think the paper indicate that the input of stacknetpose is two image patch and their seven body part detection result. Well, author did refer stacknet as network which input is two image stack. I found the repo is quite computation infeasible, using dmax=100 will cost too much time. And deepmatching is quite show too. After long time, MOT16-02 calculate 30 million cost and output edge.txt and liftedge.txt, but running c++ code for graph solver result in "cannot relocate memory". Any idea?

chenyuan1994 commented 6 years ago

@shensheng27 Sorry, I ignored the input of the stacknet in the repo . the repo do not include pose estimation net as I know. the result of the repo is not match the result in paper in my practice. The deepmatching has GPU version . it is faster but it also cost much time because there are too much image pair. it is available here. http://lear.inrialpes.fr/src/deepmatching/ And I had meet the problem of the shortage of memory. I found out that all the images were loaded in the memory first. So I change to load images when they needed.

shensheng27 commented 6 years ago

@chenyuan1994 Thanks for reply. Deepmatching do has a gpu version which is quite hard for me to configuration. I found the docker version which is useful. However, using downscale 2 one 1920*1080 image will over 8gb gpu memory, so i just take slow cpu version.

By the way, I am using poi detection which has 14092 detection on MOT16-02 and 30million edges graph when dmax=100, the graph solver run for 3days and no result yet. Is this normal?

chenyuan1994 commented 6 years ago

I have take 2 GPU for graph solver, each batch would coast 10 second. And the deepmatching has been completed prior to it. @shensheng27

shensheng27 commented 6 years ago

@chenyuan1994 Thanks for reply. It take 4 days to solve MOT16-02. I using the repo https://github.com/bjoern-andres/graph which only use cpu. So how to apply gpu, by using gurobi? Is there any guide?

MADONOKOUKI commented 6 years ago

@chenyuan1994 @shensheng27 I also want to use gpu version of deepmatch....Please tell me how to set up it!!

chenyuan1994 commented 6 years ago

@MADONOKOUKI I set up gpu version of deepmatching follow the step in https://github.com/lpplbiubiubiub/DeepMatchingGPU

chenyuan1994 commented 6 years ago

I also used cpu for graph Algorithms, but it took not much time as you spent, maybe 8 hours for MOT16-11. @shensheng27

MADONOKOUKI commented 6 years ago

@chenyuan1994 thank you!!

duyhominhnguyen commented 4 years ago

@chenyuan1994 @shensheng27 Can you download the model .h5 file in line 56 in MultiplePeopleTracking.py: "reid = StackNet64x64(data_root)"?