Inference on custom dataset ( point cloud file in the same format of SemanticKitty (.bin)

[DavideCoppola97]

Hello and thanks for your work!
I generated my dataset using a simulator, trying to create point clouds as similar as possible to those of SemanticKitty, then with a 64 beam velodyne and trying to use the same fov parameters and number of points.
The problem is that I only have the .bin files but I don't have the labels, I would simply like to use your network to make inference on my points and generate the labels, what can I do? Thank you

[edwardzhou130]

You should be able to generate the label predictions using the same testing script if you change the test data path to your custom dataset:

PolarSeg/test_pretrain_SemanticKITTI.py

Lines 133 to 165 in 3531f15

	# test
	print('*'*80)
	print('Generate predictions for test split')
	print('*'*80)
	pbar = tqdm(total=len(test_dataset_loader))
	with torch.no_grad():
	for i_iter_test,(_,_,test_grid,_,test_pt_fea,test_index) in enumerate(test_dataset_loader):
	# predict
	test_pt_fea_ten = [torch.from_numpy(i).type(torch.FloatTensor).to(pytorch_device) for i in test_pt_fea]
	test_grid_ten = [torch.from_numpy(i[:,:2]).to(pytorch_device) for i in test_grid]
	predict_labels = my_model(test_pt_fea_ten,test_grid_ten)
	predict_labels = torch.argmax(predict_labels,1).type(torch.uint8)
	predict_labels = predict_labels.cpu().detach().numpy()
	# write to label file
	for count,i_test_grid in enumerate(test_grid):
	test_pred_label = predict_labels[count,test_grid[count][:,0],test_grid[count][:,1],test_grid[count][:,2]]
	test_pred_label = train2SemKITTI(test_pred_label)
	test_pred_label = np.expand_dims(test_pred_label,axis=1)
	save_dir = test_pt_dataset.im_idx[test_index[count]]
	_,dir2 = save_dir.split('/sequences/',1)
	new_save_dir = output_path + '/sequences/' +dir2.replace('velodyne','predictions')[:-3]+'label'
	if not os.path.exists(os.path.dirname(new_save_dir)):
	try:
	os.makedirs(os.path.dirname(new_save_dir))
	except OSError as exc:
	if exc.errno != errno.EEXIST:
	raise
	test_pred_label = test_pred_label.astype(np.uint32)
	test_pred_label.tofile(new_save_dir)
	pbar.update(1)
	del test_grid,test_pt_fea,test_index
	pbar.close()

However, the quality of the prediction might not be very good. I suggest you can first visualize 5-10 random predictions to see if it meets your need.

[DavideCoppola97]

You should be able to generate the label predictions using the same testing script if you change the test data path to your custom dataset:

PolarSeg/test_pretrain_SemanticKITTI.py

Lines 133 to 165 in 3531f15

	# test
	print('*'*80)
	print('Generate predictions for test split')
	print('*'*80)
	pbar = tqdm(total=len(test_dataset_loader))
	with torch.no_grad():
	for i_iter_test,(_,_,test_grid,_,test_pt_fea,test_index) in enumerate(test_dataset_loader):
	# predict
	test_pt_fea_ten = [torch.from_numpy(i).type(torch.FloatTensor).to(pytorch_device) for i in test_pt_fea]
	test_grid_ten = [torch.from_numpy(i[:,:2]).to(pytorch_device) for i in test_grid]
	predict_labels = my_model(test_pt_fea_ten,test_grid_ten)
	predict_labels = torch.argmax(predict_labels,1).type(torch.uint8)
	predict_labels = predict_labels.cpu().detach().numpy()
	# write to label file
	for count,i_test_grid in enumerate(test_grid):
	test_pred_label = predict_labels[count,test_grid[count][:,0],test_grid[count][:,1],test_grid[count][:,2]]
	test_pred_label = train2SemKITTI(test_pred_label)
	test_pred_label = np.expand_dims(test_pred_label,axis=1)
	save_dir = test_pt_dataset.im_idx[test_index[count]]
	_,dir2 = save_dir.split('/sequences/',1)
	new_save_dir = output_path + '/sequences/' +dir2.replace('velodyne','predictions')[:-3]+'label'
	if not os.path.exists(os.path.dirname(new_save_dir)):
	try:
	os.makedirs(os.path.dirname(new_save_dir))
	except OSError as exc:
	if exc.errno != errno.EEXIST:
	raise
	test_pred_label = test_pred_label.astype(np.uint32)
	test_pred_label.tofile(new_save_dir)
	pbar.update(1)
	del test_grid,test_pt_fea,test_index
	pbar.close()

However, the quality of the prediction might not be very good. I suggest you can first visualize 5-10 random predictions to see if it meets your need.

Thanks even if I was already able to make the inference, both on this and on other networks (RangeNet ++, SqueeSegV3 etc.). Unfortunately the results are really poor, probably due to the fact that my data are obtained from a simulator (Carla) which in turn simulates a lidar, the networks cannot map correctly even the road, as soon as available I will try to do the end tuning to try to improve the performance of the networks on my pointclouds