Function fetch_unique_view_index

dataset_loaders/load_7Scenes.py:346–390

local NeRF data selection

(args, train_set, threshold)

    return train_set, val_set, bounds

def fetch_unique_view_index(args, train_set, threshold):
    ''' local NeRF data selection '''
    # Frustum near-far threshold
    K_APPROX, FRUSTUM_APPROX, W, H = camera_frustum_initializer(args)

    # compute how many unique views in the train set
    keyframe_idx = 0
    unique_frame_index = []
    unique_frame_index.append(keyframe_idx)

    for i in range(len(train_set)):
        if i % 200 == 0:
            print ('Image {:d} / {:d}'.format(i, len(train_set)))
        
        if keyframe_idx == i:
            continue
        
        # compute frustum overlap
        overlap_1p = compute_frustums_overlap(
            train_set.poses[keyframe_idx].reshape(3,4), train_set.poses[i].reshape(3,4), FRUSTUM_APPROX, K_APPROX, W, H
        )
        if overlap_1p > threshold:
            continue
            
        # double check with existing unique views
        save_flag = True
        for j in unique_frame_index:
            overlap_2p = compute_frustums_overlap(train_set.poses[j].reshape(3,4), train_set.poses[i].reshape(3,4), FRUSTUM_APPROX, K_APPROX, W, H)
            if overlap_2p > threshold:
                save_flag = False
                break

        if not save_flag:
            print('image[{}] is overlapped with unique_frame_index {}'.format(i, j))
            keyframe_idx = i
            continue

        unique_frame_index.append(i)
        print(unique_frame_index)
        keyframe_idx = i
    unique_frame_index = np.array(unique_frame_index)
    print('total train set selected', len(unique_frame_index))
    save_unique_view_dir = osp.join(args.datadir, 'unique_view.txt')
    np.savetxt(save_unique_view_dir, unique_frame_index, fmt="%d")
    return unique_frame_index

def select_nearest_neighbor_views(args, train_set, threshold, unique_frame_index):
    ''' select nearest_neighbor views based on unique_frame_index  '''