Hello authors, thank you for this great work!

In my understanding, during the training period, instances with novel classes are also exposed to the network, so the queries of these novel instances will match the "no-object" textual embeddings by the cross entropy loss. In this case, the model may tend to ignore the novel instances during inference and cause severe performance drop. But it seems not a problem in the experiements, so I'm wondering if my understanding has something wrong.

Hello authors, thank you for this great work!

I was trying to run Zegformer on Pascal VOC, but was facing some problems. I was wondering if you can release the configs and data preprocessing required to reproduce the results on this dataset reported in the paper. Would help a lot!

Thank you.

Hi, Thank you for the amazing work, Is there a way to get predictions on a new test image without downloading the whole datsets ? e.g when I run python3 demo/demo.py --config-file configs/coco-stuff/zegformer_R101_bs32_60k_vit16_coco-stuff_gzss_eval.yaml --input figures/dumy.png --output figures/dumdum.png --opts MODEL.WEIGHTS checkpoints/zegformer_R101_bs32_60k_vit16_coco-stuff.pt I get the error message FileNotFoundError: [Errno 2] No such file or directory: 'datasets/coco/coco_stuff/split/seen_classnames.json'

Can I have these seen_classnames.json files, like the .npy files you have uploaded? Or Can you explain the format of expected .json files so I can generate myself?

Thanks for the good research.

I have a question about your paper. As far as I know, there are inductive settings and transductive settings for the zero-shot task. In the zero-shot classification task, the inductive setting can be clearly implemented, but in the ZS3 (zero-shot semantic segmentation) task, in the training stage, I know that the part corresponding to the unseen class in the input image is used without masking.

So I'm confused whether the ZegFormer is inductive or transductive setting. In the case of zegformer, does it belong to inductive setting in ZS3 task?

Hello, author, you directly use CLIP model to classify the class-agnostic binary mask during the testing phase. This seems to violate the principle of zero-shot learning, because CLIP already has the information of unseen classes.

when i run the command of "python datasets/coco-stuff/prepare_coco_stuff_sem_seg_seen.py", i met the error of "No such file or directory: 'datasets/coco/coco_stuff/annotations/val2017_seen'".

Hello author, I didn't find the value of "cfg.MODEL.MASK_FORMER.PIXEL_DECODER_NAME" when checking the configuration file, because it's the first time I read the code related to detectron2, so I don't know it very well.Can you give me some help?

@dingjiansw101 Hi Jian, thanks for your great work! I am wondering did you happen to test your trained coco-stuff model directly on the ADE-20K dataset? Because in the concurrent works, like [1][2], they all report this transfer number. It is very interesting to compare your work with counterparts. Thanks!

[1] Xu, Mengde, et al. "A simple baseline for zero-shot semantic segmentation with pre-trained vision-language model." arXiv preprint arXiv:2112.14757 (2021). [2] Ghiasi, Golnaz, et al. "Open-vocabulary image segmentation." arXiv preprint arXiv:2112.12143 (2021).