Yuqian Yuan

@Unrealluver Yes, the CLIP-guided semantic segmentation is genuinely free from any annotations, following the setting in MaskCLIP+. Its robust performance can be attributed to the priori knowledge of CLIP's exceptional...

@yahooo-m Thanks, we will release the data annotation pipeline later.

Hi, @yahooo-m @xushilin1 @xin-ran-w Thanks for your attention, we have been busy with some other things recently, not having a lot of time to organize the code. While the preliminary...

Hi, @koda-11. You can add a line `classes = classes.split(‘,’)[0]` at line 74, where the error is raised, in `detectron2/detectron2/evaluation/cityscapes_evaluation.py` and it works.

你好， @DimplesL image token个数为1024，而每个区域对应的mask token和position token分别都只有1个，详见https://github.com/CircleRadon/Osprey/blob/ca9f26dbd9a0907d8ff686588a394fa897b60828/osprey/model/osprey_arch.py#L184-L187 正常训练和推理一般是不会超过2048的。

Hi, @Glupapa The starting checkpoint is our final model `Osprey-7b`, and the prompt template is the same as RefCOCO as in https://github.com/CircleRadon/Osprey/blob/ca9f26dbd9a0907d8ff686588a394fa897b60828/osprey/datasets/stage2_data.py#L256C26-L262C1 The config is as follows, ```python #!/bin/bash export...

@YuWeigang It's right.

Yes, we also have encountered this issue, because we add the drawing bounding box function from the official gradio, so `share=True` may not support. However, there is no need to...

Hi, @Glupapa For open-vocabulary segmentation, all approaches employ ground truth boxes/masks as input to assess regional recognition capability. We leverage the semantic similarity as the matching measurement to calculate these...

@Glupapa The groundtruth masks are used in calculating these metrics.