haibo-qiu / gfnet Goto Github PK

I use GFnet to train my own dataset, and the performance of the RV branch is very low. I canceled all data augmentation operations. The miou of RV is only about 10, have you ever encountered this problem?

How much memory is needed for inference with nuscenes? I have 70G of free memory, but I found that the memory was fully occupied and the system hanged. What caused the excessive memory usage? I didn't have this problem during training.

Hi. Thanks for your open source work. I have some ques about details.

1. Look like you only use trans Data Aug in Train. And look like the data aug code is similar with salsanext.
   As far as I know, kprnet's special unfolding projection method, also called proj_version='v2' in your code, causes out-of-bounds errors when using data aug, which I think is the reason why you only use one type of aug.
   However, in my past experiments, errors occurred even when using only similar trans Data Aug.
   So I want to know, did the data augmentation work fine in your experiment?
   

   GFNet/configs/resnet_semantickitti.yaml

   Lines 65 to 72 in e2e8ede

   	range:
   	max_points: 130000 # max of any scan in dataset
   	type: "spherical" # projective
   	proj: "v2" # ['v1', 'v2']
   	hres: True # good
   	flip: False # bad
   	trans: True # good
   	rot: False # bad

   
   

   GFNet/libs/utils/laserscan.py

   Lines 103 to 114 in e2e8ede

   	if whether_aug(self.train):
   	if whether_aug(self.train, self.flip):
   	self.points[:, 1] = -self.points[:, 1]
   	if whether_aug(self.train, self.trans): # should be related to the scale of points coord
   	jitter_x = random.uniform(-5, 5)
   	jitter_y = random.uniform(-3, 3)
   	jitter_z = random.uniform(-1, 1)
   	self.points[:, 0] += jitter_x
   	self.points[:, 1] += jitter_y
   	self.points[:, 2] += jitter_z
   	if whether_aug(self.train, self.rot):
   	self.points = self.points @ R.random(random_state=1234).as_dcm().T

2. Have you studied the effect of different choices of K by KNN in kpconv? Your implementation chooses the same 7 as KPRNet, and I'm wondering if this affects performance.
   

   GFNet/libs/dataloader/SemanticKitti.py

   Lines 395 to 397 in ffafd0a

   	except:
   	tree = kdtree(scan[:, :3])
   	_, knns = tree.query(scan[:, :3], k=self.neighbor)

Hi @haibo-qiu, thank you for open-sourcing your work!

Regarding the experimental results of GFNet, could you share with me what score you got on the val set of nuScenes? Thanks!

can u push it on dockerhub? QAQ

Hi, thanks for your great work and share the code!
When I try to reproduce your result, the error occurs because the knn fiiles don't exist. I would like to know how to generate knn files.