oafolabi
commented
6 years ago init is the constructor for the class. According to https://micropyramid.com/blog/understand-self-and-__init__-method-in-python-class/ :
"init" is a reseved method in python classes. It is known as a constructor in object oriented concepts. This method called when an object is created from the class and it allow the class to initialize the attributes of a class. It gets called when you do create an object in that class, so in our case when you do something like: https://github.com/charlesq34/pointnet2/blob/7961e26e31d0ba5a72020635cee03aac5d0e754a/scannet/train.py#L72
It pretty much just reads the data file to get the scene points and labels. It also creates a histogram for all the classes telling us the number of points in this class. This histogram is used to create weights that may be used to ameliorate class inbalance issues during training by weighing the loss per class.
get item is also another pyhton function used to index elements. I'm not a python expert, but I believe it is called when we try to access an element of the class, jus tlike an array. So for example, when we do something like https://github.com/charlesq34/pointnet2/blob/7961e26e31d0ba5a72020635cee03aac5d0e754a/scannet/train.py#L358 i.e. we intend to get the data at index "batch_index".
In this case, the code itself breaks us the scene at the desired index into volumes of 1.5x1.5 in the x-y plane. Points are then sampled from each grid and the collection of all the grids is returned. Please refer to the Scannet experiment details in the supplementary part of the paper for more information. This is where 1.5 comes from.
As to where 0.2 comes from, that may have just been based on trial and error/ intuition to see what works best.
Hopefully that helps. Please remember to close the issue if that solves your problem.
Coastline2018
oaa3wf
XueDoudou
In the data preparation stage, what does the bold area do and where are those parameter values come from (1.5, 0.2, etc) ?
Thank you very much for the help!
class ScannetDatasetWholeScene(): def init(self, root, npoints=8192, split='train'): self.npoints = npoints self.root = root self.split = split self.datafilename = os.path.join(self.root, 'scannet%s.pickle'%(split)) with open(self.data_filename,'rb') as fp: self.scene_points_list = pickle.load(fp) self.semantic_labels_list = pickle.load(fp) if split=='train': labelweights = np.zeros(21) for seg in self.semantic_labelslist: tmp, = np.histogram(seg,range(22)) labelweights += tmp labelweights = labelweights.astype(np.float32) labelweights = labelweights/np.sum(labelweights) self.labelweights = 1/np.log(1.2+labelweights) elif split=='test': self.labelweights = np.ones(21) def getitem(self, index): point_set_ini = self.scene_points_list[index] semantic_seg_ini = self.semantic_labels_list[index].astype(np.int32) coordmax = np.max(point_set_ini,axis=0) coordmin = np.min(point_set_ini,axis=0) *nsubvolume_x = np.ceil((coordmax[0]-coordmin[0])/1.5).astype(np.int32) nsubvolume_y = np.ceil((coordmax[1]-coordmin[1])/1.5).astype(np.int32) point_sets = list() semantic_segs = list() sample_weights = list() isvalid = False for i in range(nsubvolume_x): for j in range(nsubvolume_y): curmin = coordmin+[i1.5,j1.5,0] curmax = coordmin+[(i+1)1.5,(j+1)1.5,coordmax[2]-coordmin[2]] curchoice = np.sum((point_set_ini>=(curmin-0.2))(point_set_ini<=(curmax+0.2)),axis=1)==3 cur_point_set = point_set_ini[curchoice,:] cur_semantic_seg = semantic_seg_ini[curchoice] if len(cur_semantic_seg)==0: continue mask = np.sum((cur_point_set>=(curmin-0.001))(cur_point_set<=(curmax+0.001)),axis=1)==3 choice = np.random.choice(len(cur_semantic_seg), self.npoints, replace=True) point_set = cur_point_set[choice,:] # Nx3 semantic_seg = cur_semantic_seg[choice] # N mask = mask[choice] if sum(mask)/float(len(mask))<0.01: continue sample_weight = self.labelweights[semantic_seg] sample_weight = mask # N point_sets.append(np.expand_dims(point_set,0)) # 1xNx3 semantic_segs.append(np.expand_dims(semantic_seg,0)) # 1xN sample_weights.append(np.expand_dims(sample_weight,0)) # 1xN point_sets = np.concatenate(tuple(point_sets),axis=0)**