search

mli0603 / stereo-transformer

Revisiting Stereo Depth Estimation From a Sequence-to-Sequence Perspective with Transformers. (ICCV 2021 Oral)
Apache License 2.0
659 stars 106 forks source link

question about reproducing results on sceneflow #76

Open Torment123 opened 1 year ago

Torment123 commented 1 year ago

Hi, I'm very interested in your work. And thanks for the detailed document. Currently I'm trying to train STTR from scratch on the sceneflow dataset. As mentioned in the document, there is only a subset of sceneflow dataset has the groundtruth of occlusion (a subset of Flyingthings3d). So my understanding is that that 1. when training on samples within this subset, the forward process follows the complete procedure, including occlusion inference and occlusion loss update 2. when training on samples outside of this subset, the forward process does not include the occlusion inference and occlusion loss part. 3. After the training, the EPE is evaluated at the entire testing set of sceneflow, while the occlusion IOU is evaluated at the test set of the subset ? Sorry for the wall of text, I'd like to make sure if there is any misunderstanding if I want to reproduce the results, thanks in advance

mli0603 commented 1 year ago

Hi @Torment123

Thanks for reaching out.

1: Yes, occlusion information is used to train STTR

2-3: I have not trained on dataset without occlusion information. The scene flow result is reported on FlyingThings3D only following all prior work. Other subsets are too small and simple to be useful. I guess I used the term FlyingThings3D and SceneFlow interchangeably. Sorry for the confusion.

Torment123 commented 1 year ago

I see, thanks for the clarification.