search

lmb-freiburg / flownet2

FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
https://lmb.informatik.uni-freiburg.de/Publications/2017/IMKDB17/
Other
1k stars 318 forks source link

Problem about finetuning on FlyingThings3D with network “FlowNetS“ #90

Closed shidahu closed 6 years ago

shidahu commented 7 years ago

Hi, I try to reproduce your experimetal results in Tabel 1 given in the paper.

Step-1: I train FlowNetS with Solver-long using FlyingChairs train&test dataset, then I evaluate the trained model on Sintel train clean dataset, the EPE value is 4.28 which is close to 4.24 given in paper.

step-2: I finetune the pre-trained model in step-1 with Solver-fine using FlyingThing3D train dataset (I used flow files in _intofuture/left and the corresponding image pairs, totally 20151 trituples). However, the EPE value is 3.98 which is greater than 3.79 given in paper with a large margin.

So, is there any problem in my experiment? Should I use FlyingThing3D train&test dataset to finetune the model? When finetune the model, I set _batchsize=4 with _cropsize 768x384 while the _batchsize in FlowNet2-S_train.prototxt.template is 8, whether I use correct parameters or not?

nikolausmayer commented 6 years ago

Hi, step 1 sounds good. What is your specific training set for step 2? 20151 samples is not the entire training set, right? Good question about the batchsize. I'm afraid I'm not sure about FlowNet2-S and how its batchsize depends on the data size... please contact Eddy Ilg about this (https://lmb.informatik.uni-freiburg.de/people/ilge/)!

shidahu commented 6 years ago

@nikolausmayer , I use FlyingThing3D train dataset (folder _intofuture/left) in step 2, it contains 22390 images, but the last image in each sub-folder cannot be used. Thus, we can only generate 20151 image pairs. In the past few days, I run step 2 again. This time I use all FlyingThing3D dataset (totally 24048 samples) with _batchsize=8, the EPE evaluated on Sintel train clean dataset is 3.90, which is close to 3.79 given in the paper.

nikolausmayer commented 6 years ago

Ah yes, of course. I forgot and just counted the frames... sounds good!

zmlshiwo commented 6 years ago

@shidahu How do you generate train/test (folder into_future/left) list? Which script is used?

shidahu commented 6 years ago

@zmlshiwo I write a python script to scan the folder and generate train/test list.

zmlshiwo commented 6 years ago

@shidahu I try to generate the list, but failed. Could you send me your python script about generate flyingthings3D train/test list? Thank you. My email is "zmlshiwo@outlook.com"

nikolausmayer commented 6 years ago

(closed due to inactivity)

zhanghaozhou commented 5 years ago

@zmlshiwo Hi zmlshiwo,

I am also trying to reproduce the results in the paper. Since the script for evaluating the EPE is not provided, could you please tell me how you evaluated your models?

Thanks a lot!

Haozhou

zhanghaozhou commented 5 years ago

@shidahu Hi shidahu, I am also trying to reproduce the results in the paper. Since the script for evaluating the EPE is not provided, could you please tell me how you evaluated your models?

Thanks a lot!

Haozhou