Where to obtain the flat_inputs for S3DIS dataset?

[Xiaobing Han]

Hi Authors:

Where to obtain the   self.inputs['neighbor_last'] = flat_inputs[4 * num_layers + 4]    in line 40 of RandLANet.py? And, where to obtain the flat_inputs in the RandLANet code? Thanks.
Best,

Xiaobing

[Tao Lu]

Hi, Xiaobing. The related codes are posted as follows,

CGA-Net/examples/RandLA-Net/main_S3DIS.py

Lines 178 to 191 in e7a46cb

 neighbour_last = tf.py_func(DP.knn_search, [batch_xyz, batch_xyz, cfg.k_n], tf.int32) 
 for i in range(cfg.num_layers): 
     neighbour_idx = tf.py_func(DP.knn_search, [batch_xyz, batch_xyz, cfg.k_n], tf.int32) 
     sub_points = batch_xyz[:, :tf.shape(batch_xyz)[1] // cfg.sub_sampling_ratio[i], :] 
     pool_i = neighbour_idx[:, :tf.shape(batch_xyz)[1] // cfg.sub_sampling_ratio[i], :] 
     up_i = tf.py_func(DP.knn_search, [sub_points, batch_xyz, 1], tf.int32) 
     input_points.append(batch_xyz) 
     input_neighbors.append(neighbour_idx) 
     input_pools.append(pool_i) 
     input_up_samples.append(up_i) 
     batch_xyz = sub_points 
  
 input_list = input_points + input_neighbors + input_pools + input_up_samples 
 input_list += [batch_features, batch_labels, batch_pc_idx, batch_cloud_idx, neighbour_last] 

[Xiaobing Han]

@inspirelt
Hi，
I have one question. The code you released with CGA on S3DIS dataset, what the result is? Is the mIoU result: 65.4?

[Tao Lu]

@whuhxb
Hi, the result of the code is 68.6% on Area 5 of S3DIS.

[Xiaobing Han]

@inspirelt OK. I see. Thanks for your reply. The result 68.6% on Area 5 of S3DIS is using CloserLook3D backbone.

[Xiaobing Han]

@inspirelt Hi, I found that I can't obtain the results on Area 5 of S3DIS using RandLANet reported in your paper, the improvement is smaller than your paper. I have no idea what the problem is.

[Tao Lu]

Have you changed any experiment settings? If using the default settings, you may try to use some other checkpoint files (I mainly used the last two checkpoint files) for testing.

[Xiaobing Han]

@inspirelt
I used the default experiment settings. Using which checkpoint files for testing? Do you have any suggestion?

[Xiaobing Han]

@inspirelt Hi，how to calculate the improved area as shown in Figure 5?

[Tao Lu]

Sorry for the late reply. I use the last two saved checkpoint files for testing. The improved area is calculated by

right_origin = (pred_origin==gt_label)
right_cga = (pred_cga==gt_label)
improved_area = (right_origin==False and right_cga==True)