Regarding the experimental results on the nusc dataset

[ALLLLecy]

I found that in your code, you didn't use the transfusion detector

FSHNet/pcdet/models/dense_heads/transfusion_head.py

Line 195 in cf70535

# heatmap = heatmap * (heatmap == local_max)

Is this correct?

[Say2L]

We modified the post-process operation in the transfusion head. The maxpool operation is replaced with NMS.

[ALLLLecy]

Thank you very much for your answer. Since I don't have sufficient experimental resources, I'm unable to reproduce which of the two effects is better.
May I ask if the indicators would be better under this experimental configuration?
Furthermore, I noticed that in the nusc detector, you used the heatmap constructed by conv2d.

FSHNet/pcdet/models/dense_heads/transfusion_head.py

Lines 105 to 110 in cf70535

	self.shared_conv = nn.Conv2d(in_channels=input_channels,out_channels=hidden_channel,kernel_size=3,padding=1)
	layers = []
	layers.append(BasicBlock2D(hidden_channel,hidden_channel, kernel_size=3,padding=1,bias=bias))
	layers.append(nn.Conv2d(in_channels=hidden_channel,out_channels=num_class,kernel_size=3,padding=1))
	self.heatmap_head = nn.Sequential(*layers)
	self.class_encoding = nn.Conv1d(num_class, hidden_channel, 1)

On the original configuration of SAFDNet, subconv2d is used here. Would this setting be better?
https://github.com/zhanggang001/HEDNet/blob/8694e7f07e860416bd99dbf46f952b3e6b6c0c6f/pcdet/models/dense_heads/sparse_transfusion_head.py#L123-L130

[Say2L]

In our experiments, the current post-process strategy performed slightly better than the original one. The two types of convolutions performed very closely in our method.