YUAN, Zhihao

icon indicating a website link https://blog.csdn.net/happy990 icon indicating an email [email protected]

icon company The Chinese University of Hong Kong, Shenzhen icon location Shenzhen, China icon location Ph.D student at CUHK(SZ). Focus: 3D computer vision and language.

Results 13 comments of YUAN, Zhihao

About co-attention problem

Hi @sH1cHEnG , You can refer to https://github.com/MILVLG/mcan-vqa for the implementaion. We may release the code in the future. Best, Zhihao

About evaluating on Referit3D

Hi @ayushjain1144 , the Acc@mIoU is only used in ScanRefer. For referit3d, we use GT instance as input and evaluate the grounding accuracy solely. Please refer to their paper for...

conv3d with empty kernel_map

Same question here. Do you have the plan to support directly inverse conv3d?

Cannot reproduce the PointGroup Detector performance

Yes, the checkpoint can get satisfactory results. As I said above, I can get the detection result mAP@50 around 50 with the given checkpoint.

Add the CIDEr metric?

Any progress?

Dimension mismatch while loading model from checkpoint

Delete `[:self.max_des_len]` here. https://github.com/daveredrum/D3Net/blob/b505e984cc4b01ea6ed95aa94b7bafa45215f4f4/lib/dataset/pipeline.py#L453

Can't reproduce the performance on real_example charlie.png

Same problem here.

Some questions about the code

X-Trans2Cap model use 2D information in training stage only.

Some questions about the code

You can use the teacher model in inference. Uncomment below and change input feats accordingly. https://github.com/CurryYuan/X-Trans2Cap/blob/c78a27209f14fcbbec74fe8b5edc06faea2e7d44/models/xtrans.py#L302

nr3d dataset

Please save the nr3d.csv to json file similar to scanrefer.