Function eval_queue

unipixel/eval/eval_groundmore.py:242–337

(q, rank, out_dict, pred_path, video_root)

def eval_queue(q, rank, out_dict, pred_path, video_root):
    while not q.empty():
        video, exp_idx = q.get()

        expressions = metadata[video]["questions"]
        expression_list = list(expressions.keys())

        clip_start = video[-9:].split("_")[0][:2] + ":" + video[-9:].split("_")[0][2:]
        # clip_end = video[-9:].split("_")[1][:2] + ":" + video[-9:].split("_")[1][2:]

        # read all the anno meta
        meta = {}
        meta["video_id"] = video
        meta["exp"] = expressions[expression_list[exp_idx]]["question"]
        meta["ans"] = expressions[expression_list[exp_idx]]["answer"]
        meta["obj_id"] = int(expressions[expression_list[exp_idx]]["obj_id"])
        meta["q_type"] = expressions[expression_list[exp_idx]]["q_type"]
        meta["exp_id"] = expression_list[exp_idx]

        start = expressions[expression_list[exp_idx]]["action_start"]
        end = expressions[expression_list[exp_idx]]["action_end"]
        action_start = (time_str_to_seconds(start) - time_str_to_seconds(clip_start)) * 6  # fps=6
        action_end = (time_str_to_seconds(end) - time_str_to_seconds(clip_start)) * 6 - 1

        meta["action_start"] = action_start
        meta["action_end"] = action_end
        # meta["frame_dir"] = frame_start.zfill(4) + "_" + frame_end.zfill(4)

        # 2. For each expression
        video_id = meta["video_id"]
        exp_id = meta["exp_id"]
        obj_id = meta["obj_id"]
        q_type = meta["q_type"]

        # action start and end is used to obtain gt masks in temporal dimension
        action_start = meta["action_start"]
        action_end = meta["action_end"]

        frame_dir = os.path.join(video_root, video_id, "images/")
        if not os.path.exists(frame_dir):
            print("Missing frames: {}.".format(video_id))
            continue
        raw_frames = nncore.ls(frame_dir, ext='jpg')  # all the frames
        has_frm = nncore.pure_name(raw_frames[0]).startswith('frame_')
        sample_indices = np.linspace(0, len(raw_frames) - 1, num=20, dtype=int)  # uniformly sample 20 frames
        assert len(sample_indices) == 20

        preds = nncore.ls(f'{args.pred_path}/{video_id}/{exp_id}', ext='png', join_path=True)
        preds.sort(key=lambda p: int(re.sub(r'^\D*', '', nncore.pure_name(p))))
        # assert len(preds) == len(sample_indices), (f'{video_id}/{exp_id}', len(preds), len(sample_indices))

        pred_0 = cv2.imread(preds[0], cv2.IMREAD_GRAYSCALE)
        h, w = pred_0.shape
        origin_h, origin_w = pred_0.shape
        all_pred_masks = np.zeros((len(sample_indices), h, w), dtype=np.uint8)

        for frame_idx, index in enumerate(sample_indices):
            mask_id = "frame_" + str(index).zfill(6) + ".png" if has_frm else str(index).zfill(7) + ".png"