Renrui Zhang comments

Results 56 comments of


                                            Renrui Zhang

How to use Depth embedding.

Same question. The paper said that the depth maps are transformed into disparity maps. Will this matter? @softmurata

The exact clip adapter few shot acc for 11 datasets

@gordonhu608 @liyaowei-stu The quantity results of CLIP-Adapter have been updated as [here](https://github.com/gaopengcuhk/Tip-Adapter/blob/main/exp.log) as `CLIP-A`.

The exact clip adapter few shot acc for 11 datasets

@June01 Thanks for this question. We follow the code for data pre-processing in [CoOp](https://github.com/KaiyangZhou/CoOp). Their reproduced zero-shot CLIP might have differences to the original CLIP paper.

"Tip-Adapter/main.py" use test features to eval

Thanks for your interest. Here in line 111, we conduct inference on the test set after every epoch's training.

"Tip-Adapter/main.py" use test features to eval

@waleedgondal Thanks for pointing out. Sorry for this mistake. The model selection after every-epoch training should be rectified by using validation set. By our experiments, this would not affect the...

Sharpness & residual ratio range

Hi Thanks for your interest. The alpha and beta are both set to 1 as the tuning baseline. The alpha weighs the importance between CLIP-pre-trained and few-shot knowledge. If the...

Apply for the pretrained weights on other few-shot learning

Thanks for your interest! We will complete this in a few days.

How to adapt PointCLIP for part-segmentation task

Thanks for your interest on our work. We have conducted zero-shot part segmentation on ShapeNetPart dataset in [PointCLIP V2](https://arxiv.org/abs/2211.11682) by back-projecting CLIP's prediction from multi-view images. We will soon release...

Can I know the computation?

Thanks for your interest. We only require one GPU for efficient few-shot training.

General questions

Thanks for your interests! Yes, queries will predict 3 categories according to model.num_classes, but the training is only by car samples and the car-dimension is selected as final outputs. You...