Is the default flownet2 only trained on FlyingThings3D?

[achalddave]

I'm pretty sure this is the case, but there doesn't seem to be a specific documentation.

The download_models.sh script downloads a bunch of models. I believe the default FlowNet2/FlowNet2_weights.caffemodel.h5 model is trained on FlyingThings3D only; is that correct?

[nikolausmayer]

FlowNet2 is a composite network assembled from

• FlowNet2-CSS-ft-sd, trained on
  • FlyingChairs
  • FlyingThings3D
  • ChairsSDHom
• FlowNet2-SD, trained on
  • ChairsSDHom
• and a small untrained fusion network.

I believe (but am not sure) that the complete FlowNet2 assembly was then again trained to convergence on FlyingThing3D + ChairsSDHom.

For more details, please see sections 3, 4, and 5 of the FlowNet2 paper.

[FUZhanhong]

hello sir,

what about DispNetCorr1D? trained on FlyingThings3D and finetuned with KITTI2015? the paper <A Large Dataset to Train Convolutional Networks or Disparity, Optical Flow, and Scene Flow Estimation> has mentioned the dataset FlyingChairs, if i pre-trained the net with FlyingChairs, can the result be better for outdoor scene?

another question: in another issue in programme "dispnet-flow-docker" which you have closed, you have mentioned that finetuned with KITTI2015 training data with GT(200), so the net in DispNetCorr1D-K is the same with DispNetCorr1D, or cutted the test part?

Thanks in advance!

[nikolausmayer]

@FUZhanhong DispNetCorr1D only uses FlyingThings3D. FlyingChairs is not a disparity dataset, so that probably won't get you very far. DispNetCorr1D-K is just DispNetCorr1D with additional finetuning on KITTI data. I don't understand your question about the "test part".

[FUZhanhong]

thanks for your reply!

for the "test part" in net, i mean those layers which occupy the structure " include { phase: TEST } " , we all know that the "testing" datasets in KITTI 2015 doesn't provide the GT, we only have 200 images with GT in "training", so, we indicate the source of data param is "../../../data/FlyingThings3D_release_TEST_lmdb" or ""../../../data/KITTI2015_TEST_lmdb" or just delete all the layers which occupy the structure above

thanks in advance!

[nikolausmayer]

We just tested on the training data during finetuning. Of course that's terrible practise considering overfitting, but it's KITTI, so... overfitting works :man_shrugging: