Frano Rajič

The `https://drive.google.com/drive/folders/1fOhEdHqrp_6D_tBsrR9hazDLYV2Sw1XC` link was taken from the official dataset website [here](https://sites.google.com/view/unidentified-video-object/dataset), but it does not seem to be accessible anymore. I think you should contact the dataset creators/maintainers for help.

The best result I got is 59.7, with a topk of 2, a finetuning learning rate of 4e-3, and 2000 finetuning epochs. This gives, for example, the predicted masks and...

However, I do get a close number when evaluating on the DAVIS 2016 (not 2017) validation subset and with hyperparameters suggested in the paper (topk=2, lr=4e-4, epochs=800), wondering if this...

Hej, in the **first** query mode, only the future time steps need to be tracked. This setting is particularly relevant in an online context. However, in the **strided** mode, all...

Hi Jenny, I get the following numbers for the DAVIS subset of TAP-Vid and the `spaT_final.pth` checkpoint shared by the authors: ```css tapvid_davis_avg_Jaccard 59.10832733877460 tapvid_davis_avg_delta 75.03060472325210 tapvid_davis_avg_OA 89.15536883348100 ``` My...

Perhaps try running with half precision? E.g., see PyTorch docs [here](https://pytorch.org/tutorials/recipes/recipes/amp_recipe.html) or use [PyTorch Lightning](https://lightning.ai/docs/pytorch/1.5.9/advanced/mixed_precision.html) to wrap your code and use their out-of-the-box mixed precision flags. If you are using...

Maybe you can try using the `cupy` version reported in the `requirements.txt`, worked fine for me: https://github.com/henry123-boy/SpaTracker/blob/4e777d14852a83954e0f45a09791cc6888643f95/requirements.txt#L10

Those keys should be there, yes. What is your `timm` version? Maybe try `pip install timm=0.6.7` as `timm` is used here: https://github.com/henry123-boy/SpaTracker/blob/4e777d14852a83954e0f45a09791cc6888643f95/models/monoD/zoeDepth/midas_c/midas/backbones/beit.py#L9

The `self.model` will output relative depth in that snippet as the `depth_anything_vits14.pth` checkpoint of DepthAnything is not a metric mono depth estimator. So the authors have used the metric depth...

Hmm, I realized that Kubric has a dome on which the floor and sky are projected and which is at a depth of around 50 Kubric units, thus the max_depth...