Bin Yan

The program may pre-allocate memory for more efficient training. More details can be found at [here](https://github.com/Megvii-BaseDetection/YOLOX/blob/57309c730d9f8c0dd1496706886ec00ddc8a11d2/yolox/utils/metric.py#L31)

Hi, "learnable broadcast sum" performs almost the same as "direct broadcast sum". Both of them perform better than "broadcast multiplication". Thanks.

Got it. We have fixed it now. Please re-check it.

Hi, bdd100k is only used for data processing and evaluation. I used python3.7 and it worked fine with bdd100k.

Hi, please refer to this guide (https://github.com/MasterBin-IIAU/Unicorn/blob/master/assets/test.md). tracker_name should be "unicorn_sot" tracker_para could be chosen from the following names (unicorn_track_large, unicorn_track_tiny, unicorn_track_tiny_rt, unicorn_track_r50)

@changsubi Hi, for now, we do not have a complete demo script for all video-level tasks. However, you can refer to this [issue](https://github.com/MasterBin-IIAU/UNINEXT/issues/8) to find a simple tutorial about demo...

@jiahui1688 Hi, for datasets of these tasks, such as (Ref-)Youtube-VOS and (Ref-)DAVIS, our code will save the mask results. You can check them under outputs dir. We will consider introducing...

Hi, ViT should be downloaded only when you want to use it as the visual encoder (visual backbone). As mentioned in the paper, we always use BERT-base as the text...

@bhack Thanks for providing these new benchmarks. We haven't evaluate our methods on them yet. However, we may consider adding corresponding results in the future.

@bhack Thank you! We add test our method on them later.