Paul Voigtlaender comments

Results 15 comments of


                                            Paul Voigtlaender

for eval

Hi xuchen, note that there is no need to change the arguments to tower_setup. Instead, please change the config (create a separate config file for eval, or override using the...

Empty detections and Assertion Error while running Forwarding and Tracking

Sorry, from this explanation I can't see what causes the problem. It could very well be some problem with paths, that some data is not found. You can try to...

pytorch version

Hi, please contact Jens Piekenbrinck, he already created such a version. Maybe he's willing to share it with you. [email protected] (replace with the name)

pytorch version

I'm sorry, I confused something. Above I wrote that Jens has a PyTorch version, but this PyTorch version is actually for Siam R-CNN, not for TrackR-CNN. In fact, sadly we...

Is the code runtime environment Windows or ubuntu?

We only ran on Ubuntu. Maybe there are small issues on Windows, but it shouldn't be too hard to make it work.

mots_common/io.py does not support folder structure for MOTS dataset?

Hi, there is a slight inconsistency, we first created a version of the MOTSchallenge data in the same format as KITTI and put it on https://www.vision.rwth-aachen.de/page/mots (where you still find...

"To check that no frame contains overlapping masks" : Can we use mots_tools for overlapping masks?

Hi, what you can do is to just disable the assertion, seems you already did that. This should mostly be ok, but if masks overlap a lot then the mapping...

Incorrect seqmap for the actual train set in MOTS?

Here the same applies as in #8 There are 2 versions of the datasets with a slightly different format and the tools here currently only work with one of them.

Weak result when using my own video

I think this sequence is just extremely hard due to the high object density. You can try to tweak the hyper-parameters, especially the weighting factors of the TDPA.