train the PTv3-object model with dinov2

[TSlus]

hi, thanks for your amazing work! I have tow question :

1. in your paper you say, We first pre-train 3D backbone PTv3-object on 3D large-scale data Objaverse, distilling visual features from FeatUp-DINOv2. Do you provide this part of the training code in your code (include how to generate DINO feature)? because I want to train on custom dataset instead of using the pre-trained model directly.
2. in the part of train light-weight MLPs to distill 2D masks to scale-conditioned grouping, SAM results supervision is required. But you did not use when training Objaverse data--knight, is right ?

thanks for your reply.

[yhyang-myron]

Hi, thanks for your interest in our work.

1. The code is like this:

self.Encoder_2d = torch.hub.load("mhamilton723/FeatUp", 'dinov2').cuda().eval()

point = self.backbone(pcd_dict)
point_feat = point.feat
frames_per_mesh = input_dict["frames_per_mesh"]
imgs = input_dict["imgs"]
B = imgs.shape[0]
V = imgs.shape[1]
link = input_dict["link"]
with torch.no_grad():
img_feat_list = []
for i in range(B):
    img_f = self.Encoder_2d(imgs[i])
    img_f = F.interpolate(img_f, size=(512, 512))
    img_feat_list.append(img_f)
    img_feat = torch.stack(img_feat_list, dim=0)

frame_pcd_feat = torch.zeros_like(point_feat)
frame_pcd_feat_flag = torch.zeros((point_feat.shape[0], 1)).cuda()
for v in range(V):
    f = img_feat[link[:, 0, v], v, :, link[:, 2, v], link[:, 1, v]]
    f *= link[:, 3, v].unsqueeze(dim=1).float()
    frame_pcd_feat += f
    frame_pcd_feat_flag += link[:, 3, v].unsqueeze(dim=1)

frame_pcd_feat_mask = (frame_pcd_feat_flag > 0).squeeze(-1)
frame_pcd_feat[frame_pcd_feat_mask] = frame_pcd_feat[frame_pcd_feat_mask] / frame_pcd_feat_flag[frame_pcd_feat_mask]

loss = F.mse_loss(point_feat[frame_pcd_feat_mask], frame_pcd_feat[frame_pcd_feat_mask])

2. We use SAM results in the MLPs training stage.

[htuann2712]

Can you release full dataset class code for this distilling PTv3-object stage?
I am planning to implement your method, however I am a bit confused about the link above the code you provided.

Many thanks for your reply.

[yhyang-myron]

SAMPart3D_pretrain_dataset.txt
Hi, this is the dataset we used.