cvlab-unibo / learning2adaptforstereo Goto Github PK

Hello! Thanks for the great work! @AlessioTonioni
I'm at it again and have questions about the results of "Learning to adapt".

I used 12 Synthia video sequences as dataset and meta-trained the network you provided with the following parameters:

--dataset=./meta_datasets.csv
--batchSize=4
--weights=./pretrained_weight/weights.ckpt	# download from the link you provided
--numStep=40000
--lr=0.0001
--alpha=0.00001
--adaptationSteps=3
--metaAlgorithm=L2AWad
--unSupervisedMeta
--maskedGT

After training, I used this weight to test online adaptation on video sequences from DrivingStereo and KITTI raw data. I found that the prediction results for the first few frames were extremely poor , the error rate D1 is close to 99%, but after 100 to 200 frames, D1 quickly drops below 10%.

I would like to ask:

1. Is it normal for the initial prediction results to be so poor?
2. Is there anything wrong with my training?
3. Is this result representative of your work for comparison?

Sorry for the troublesome questions, but I'd appreciate your answers!

Thanks for you open access repo.
I would like to ask for the specific software version number,as the tensroflow 2.0 was already released recently

Dear Sir
How did you get groundtruth of KITTI datasets in the .csv file ? I have downloaded the dataset before, however there is no such file .

Hello! Thanks for the great work! @AlessioTonioni
But I have some questions about pretrained weights.

The pretrained weights you provide are exactly the same as the weights in this repository (Real-time-self-adaptive-deep-stereo).
So how do I get the weights pretrained on Carla or synthia (with meta learning)?

hello! thanks for the great work!
When I tried to run your code on my pc, I cannot run it directly and I have 3 questions about your work.

1. How do you organize your ' adaptation_list.csv '?
2. The order of training on Sceneflow, kitti and other datasets.
3. When you training the models, how many GPUs did you use? And their video memory?

thanks again!

hi, thanks for your work. I just wonder why you pretrain the model on F3D before L2A-training mode. And why when pretraining the L2A-training mode is not used?

I wonder how to update bn parameters during the adaptation stage, since the batchsize is 1.

thanks for the great work.
I have some questions about datasets for training.

1. the synthia dataset only have the depth map. when I use its provided focal length and baseline to get the disparity, it's data looks terrible. so I wanna know how you process these data.
2. I transform sparse depth info in kitti's video sequences, can I use these disparity data to train the net, for the data is really sparse when compared with the stereo datasets.
   best wishes.
   zilch

Thanks for the great work！
I want to know if the weighted mask will be updated in online adaptation phase using a small amount of gt labels？