search

megvii-research / CREStereo

Official MegEngine implementation of CREStereo(CVPR 2022 Oral).
Apache License 2.0
483 stars 59 forks source link

What datasets are used for pretraining? #5

Closed DY-ATL closed 2 years ago

DY-ATL commented 2 years ago

The pretrained model works amazingly well on the real-life photos! What datasets are used for pretraining? Can you please provide the training details of the pretrained model? Thanks!

JackLee396 commented 2 years ago

As described in Sec 4.1 in the paper, in addition to our rendered dataset, we collect major public datasets for training including Sceneflow, Falling Things, Airsim, etc. No pre-training steps are implemented.

ynma-hanvo commented 10 months ago

As described in Sec 4.1 in the paper, in addition to our rendered dataset, we collect major public datasets for training including Sceneflow, Falling Things, Airsim, etc. No pre-training steps are implemented.

hi, how did you convert the sceneflow dataset into crestereo format, as i checked its content, there is only flow data which looks different from disparity.

hgchen commented 9 months ago

A quick question about the provided pre-trained model "crestereo_eth3d.mge". We wonder what dataset that particular model was trained on. According to your paper section 4: "In addition to our rendered dataset, we collect major public datasets for training, including Sceneflow [27], Sintel [5] and Falling Things [46]." We wonder if you trained that model on all the synthetic datasets, and then fine-tuned it with the ETH3D datasets?

If that's the case, you mentioned "We train our network on the whole training set with a proportion of 2% augmented training data from ETH3D low-res two-view stereo dataset." in your paper, did you use only the very small "low-res two-view" subset, or did you use the "high-res multi-view" and "low-res multi-view" subsets as well?

Thanks a lot for your answers, really appreciate your time!