Can’t reproduce the high cosine similarity result in the paper with shared ACFNet checkpoint

[sg47]

Hi, Thank you for sharing! Follow the GETTING_STARED.md, i implemented the cosine similarity myself and checked it on the demo data with the checkpoint from https://github.com/DeepMotionAIResearch/DenseMatchingBenchmark/blob/master/configs/AcfNet/ResultOfAcfNet.md#sceneflow and most of the feature similarity are still around 0.6-0.9. Is the checkpoint provided already trained for feature consistency? And it would be appreciated if a script to check the similarity could be provided. Thanks again!

[jiaw-z]

Hi, sorry to reply late. The checkpoint is trained by authors of DenseMatchingBenchmark, so it is trained without our feature consistency constraints. We are about to release our pre-trained checkpoints later. You can train models with consistency constraints with the config file in ./configs/FCStereo/. The similarity reported in the paper is calculated on valid and non-occ pixels with pixel-wise cosine similarity. The valid masks are generated in the same way as the valid masks for contrastive loss, which can be found in./dmb/modeling/stereo/losses/contrastive_loss.py. Hope this can help you!

[mattpoggi]

We are about to release our pre-trained checkpoints later.

Hi, any news about this?

[Zhaohuai-L]

Hi, can you share your pretrained checkpoint?

[jiaw-z]

Hi, can you share your pretrained checkpoint?

We are very sorry that the sharing of checkpoints may be delayed. We used a lot of storage when experimenting with our work for cvpr2023 and accidentally deleted all our previous checkpoints. We need to retrain models to obtain checkpoints for FCStereo.