Lu Qi
Lu Qi
The KINS annotations follow the COCO format. Each mask is encoded by the RLE.
The annotation includes the occlusion mask, so you just use amodal mask instead. I just use occlusion area in occlusion classification branch to indicate whether this RoI have occlusion?
I will check it and update the annotation as soon as possible. It may be after the rebuttal of ECCV. You could use the old annotation in this week. If...
Hi, The data and code will be released in July. Thanks. Lu
Hi, Thanks for your interests. The code and datasets will be released upon the acceptance of the paper. Lu
Hi, All the released models just start with COCO pre-trained weights. And then fine-tuned on the EntitySeg training dataset and evaluated on the EntitySeg val dataset.
Hello, You can directly use the existing entity segmentation model to generate entity masks. For example, directly use the CropFormer Visualization code. https://github.com/qqlu/Entity/blob/main/Entityv2/CODE.md. This could be directly used for pseudo...
Sorry, I do not understand your requirements. The CA-SSL mainly uses the unlabeled data and its class-agnostic pesudo labels to warm-up the network whatever the framework be used.
Hi, Thanks for your interest in our paper. We will release the code before the CVPR conference. Lu
Hi We will try to convert the CropFormer to the TensorRT. However, I am not sure costing 15 minutes for 50 images is the normal speed. Is your image very...