def predict_crop_occ(self, pi, c, vol_bound=None, **kwargs):
        ''' Predict occupancy values for a crop

        Args:
            pi (dict): query points
            c (tensor): encoded feature volumes
            vol_bound (dict): volume boundary
        '''
        occ_hat = pi.new_empty((pi.shape[0]))

        if pi.shape[0] == 0:
            return occ_hat
        pi_in = pi.unsqueeze(0)
        pi_in = {'p': pi_in}
        p_n = {}
        for key in self.vol_bound['fea_type']:
            # projected coordinates normalized to the range of [0, 1]
            p_n[key] = normalize_coord(pi.clone(),
                                       vol_bound['input_vol'],
                                       plane=key).unsqueeze(0).to(self.device)
        pi_in['p_n'] = p_n

        # predict occupancy of the current crop
        with torch.no_grad():
            occ_cur = self.model.decode(pi_in, c, **kwargs).logits
        occ_hat = occ_cur.squeeze(0)

        return occ_hat
Example #2
0
    def eval_points(self, p, c=None, vol_bound=None, **kwargs):
        ''' Evaluates the occupancy values for the points.

        Args:
            p (tensor): points 
            c (tensor): encoded feature volumes
        '''
        p_split = torch.split(p, self.points_batch_size)
        occ_hats = []
        for pi in p_split:
            if self.input_type == 'pointcloud_crop':
                if self.vol_bound is not None:  # sliding-window manner
                    occ_hat = self.predict_crop_occ(
                        pi, c, vol_bound=vol_bound, **kwargs)
                    occ_hats.append(occ_hat)
                else:  # entire scene
                    pi_in = pi.unsqueeze(0).to(self.device)
                    pi_in = {'p': pi_in}
                    p_n = {}
                    for key in c.keys():
                        # normalized to the range of [0, 1]
                        p_n[key] = normalize_coord(
                            pi.clone(), self.input_vol, plane=key).unsqueeze(0).to(self.device)
                    pi_in['p_n'] = p_n
                    with torch.no_grad():
                        occ_hat = self.model.decode(pi_in, c, **kwargs).logits
                    occ_hats.append(occ_hat.squeeze(0).detach().cpu())
            else:
                pi = pi.unsqueeze(0).to(self.device)
                with torch.no_grad():
                    occ_hat = self.model.decode(pi, c, **kwargs).logits
                occ_hats.append(occ_hat.squeeze(0).detach().cpu())

        occ_hat = torch.cat(occ_hats, dim=0)
        return occ_hat
Example #3
0
    def load(self, model_path, idx, vol):
        ''' Loads the data point.

        Args:
            model_path (str): path to model
            idx (int): ID of data point
            vol (dict): precomputed volume info
        '''
        if self.multi_files is None:
            file_path = os.path.join(model_path, self.file_name)
        else:
            num = np.random.randint(self.multi_files)
            file_path = os.path.join(model_path, self.file_name,
                                     '%s_%02d.npz' % (self.file_name, num))

        points_dict = np.load(file_path)
        points = points_dict['points']
        # Break symmetry if given in float16:
        if points.dtype == np.float16:
            points = points.astype(np.float32)
            points += 1e-4 * np.random.randn(*points.shape)

        occupancies = points_dict['occupancies']
        if self.unpackbits:
            occupancies = np.unpackbits(occupancies)[:points.shape[0]]
        occupancies = occupancies.astype(np.float32)

        # acquire the crop
        ind_list = []
        for i in range(3):
            ind_list.append((points[:, i] >= vol['query_vol'][0][i])
                            & (points[:, i] <= vol['query_vol'][1][i]))
        ind = ind_list[0] & ind_list[1] & ind_list[2]
        data = {
            None: points[ind],
            'occ': occupancies[ind],
        }

        if self.transform is not None:
            data = self.transform(data)

        # calculate normalized coordinate w.r.t. defined query volume
        p_n = {}
        for key in vol['plane_type']:
            # projected coordinates normalized to the range of [0, 1]
            p_n[key] = normalize_coord(data[None].copy(),
                                       vol['input_vol'],
                                       plane=key)
        data['normalized'] = p_n

        return data