How fast is your inference speed when inputting an object point cloud？

[yys-abu]

I'm not sure if your model has real-time capability, but I'm still working on the inference code. It would be great if you could provide the inference code

[hanryyan]

The paper under arxiv version (https://arxiv.org/html/2503.18007v1#Sx7) presented our running speed, which is 13.24 ms on an RTX 4090 GPU.

[yys-abu]

The paper under arxiv version (https://arxiv.org/html/2503.18007v1#Sx7) presented our running speed, which is 13.24 ms on an RTX 4090 GPU.

Thank you for your reply. I have reviewed the ShapNet55/34 dataset in PoinTr and find that it has many common objects in daily life, such as bottles which are I nedded. But when I used my own bottle point cloud to inference (using ShapNet55 weights), I found that the effect was very poor. I think it may be a problem with perspective, but there is no point cloud data in the ShapNet55 dataset you provided. So I used the airplane point cloud in the PCN dataset for testing, and (using PCN weights) I found that randomly rotating the point cloud had poor results. So I have a question, can your method generalize to common objects in daily life such as cups or bottles, and does the shooting angle of the point cloud have a significant impact on the effect. Looking forward to your reply.

[hanryyan]

I think it can be generalized to cups or bottles. However, you may need to fine-tune this model for your special requirements, for example, viewpoints. In my experiments (testing our method in real-world data), the viewpoints, scale, and rotation affect the performance of our model. In this work, for a fair comparison, we follow the setting of PoinTr to train our model on the ShapeNet55 dataset. So, we didn't use the data augmentation, such as random rotation, rescaling. If you want to improve the generalizability, you need to fine-tune the model by using the relevant data. For instance, rotating and rescaling the input point clouds. Besides, you can refer to this link (https://drive.google.com/file/d/1jUB5yD7DP97-EqqU2A9mmr61JpNwZBVK/view) to download the ShapeNet55 dataset and check the difference between your data and ShapeNet's data. It is particularly important to note that you need to normalize your input point clouds in order to achieve better results. In summary, you should first check whether your input is similar to the input used by our network. Then, fine-tune our model to adapt to personal data.

[yys-abu]

I think it can be generalized to cups or bottles. However, you may need to fine-tune this model for your special requirements, for example, viewpoints. In my experiments (testing our method in real-world data), the viewpoints, scale, and rotation affect the performance of our model. In this work, for a fair comparison, we follow the setting of PoinTr to train our model on the ShapeNet55 dataset. So, we didn't use the data augmentation, such as random rotation, rescaling. If you want to improve the generalizability, you need to fine-tune the model by using the relevant data. For instance, rotating and rescaling the input point clouds. Besides, you can refer to this link (https://drive.google.com/file/d/1jUB5yD7DP97-EqqU2A9mmr61JpNwZBVK/view) to download the ShapeNet55 dataset and check the difference between your data and ShapeNet's data. It is particularly important to note that you need to normalize your input point clouds in order to achieve better results. In summary, you should first check whether your input is similar to the input used by our network. Then, fine-tune our model to adapt to personal data.

ShapeNet55.zip contains npy files, which I converted into pcd point cloud files and found that they are all complete object point clouds, without any incomplete point clouds that need to be completed

[ZijunLi402]

ShapeNet55.zip contains npy files, which I converted into pcd point cloud files and found that they are all complete object point clouds, without any incomplete point clouds that need to be completed

Yes, the point clouds in the ShapeNet dataset are all complete. However, during model training/validation/testing, the preprocessing step misc.seprate_point_cloud(tools/runner.py line 382) is applied first, which converts the point cloud into an incomplete input before feeding it into the model.