Max Zhaoshuo Li 李赵硕

Hi @lfxx @neuzyy, The STTR builds a full resolution feature so it can sparsely sample the feature at different stride as discussed in Sec 3.5. This unfortunately, together with attention...

Hi @lfxx Yes, downsampling will definitely help. A 2x downsampling will cut the memory consumption by more than 4 times.

Hi @SpadeLiu I actually explicilty did this because if neither condition is met, there are NANs in the query and key. You can find more details in the issue #11...

Hmm interseting. I have not heard people reporting this as a random error... If you find patterns, let me know. STTR doesn't estimate the occluded areas. It sets the disparities...

All pretrained models are provided: https://github.com/mli0603/stereo-transformer#pre-trained-models The table is a reference for you to check if things on your end runs the same as mine. So if you run the...

You can finetune the sceneflow trained model (the link above) using this script: https://github.com/mli0603/stereo-transformer/blob/main/scripts/kitti_finetune.sh

Yes. The pretrained script for Scene Flow can be found here: https://github.com/mli0603/stereo-transformer/blob/main/scripts/pretrain.sh

@minchong1998 Oh I see where your confusion comes from. I used 400 epochs but I wrote 300 in the finetuning script. This is a mistake when I transferring PyCharm configuration...

Hi @minchong1998, I attched my training log at epoch 250 and 300. Can you see if they match what you have? 1. ``` Index 0, attn 0.2269, rr 0.8732, l1...

> i found the groundtruth of occ_mask you use to calculate iou is inputs.occ_mask.(self.compute_iou(outputs['occ_pred'], inputs.occ_mask, loss, invalid_mask), which is set as input_data['occ_mask'] = np.zeros_like(disp).astype(np.bool) in kitti.py. This is also use...