obj_detect_tracking.py

# coding=utf-8
"""
  run object detection and tracking inference
"""

import argparse
import cv2
import math
import json
import random
import sys
import time
import threading
import operator
import os
import pickle
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"

# remove all the annoying warnings from tf v1.10 to v1.13
import logging
logging.getLogger("tensorflow").disabled = True

import matplotlib
# avoid the warning "gdk_cursor_new_for_display:
# assertion 'GDK_IS_DISPLAY (display)' failed" with Python 3
matplotlib.use('Agg')

from tqdm import tqdm

import numpy as np
import tensorflow as tf

# detection stuff
from models import get_model
from models import resizeImage
from nn import fill_full_mask
from utils import get_op_tensor_name
from utils import parse_nvidia_smi
from utils import sec2time

import pycocotools.mask as cocomask

# tracking stuff
from deep_sort import nn_matching
from deep_sort.detection import Detection
from deep_sort.tracker import Tracker
from application_util import preprocessing
from deep_sort.utils import create_obj_infos,linear_inter_bbox,filter_short_objs

# for mask
import pycocotools.mask as cocomask

# class ids stuff
from class_ids import targetClass2id_new_nopo
from class_ids import coco_obj_class_to_id
from class_ids import coco_obj_id_to_class
from class_ids import coco_obj_to_actev_obj
from class_ids import coco_id_mapping

targetClass2id = targetClass2id_new_nopo
targetid2class = {targetClass2id[one]: one for one in targetClass2id}

def get_args():
  """Parse arguments and intialize some hyper-params."""
  global targetClass2id, targetid2class
  parser = argparse.ArgumentParser()

  parser.add_argument("--video_dir", default=None)
  parser.add_argument("--video_lst_file", default=None,
                      help="video_file_path = os.path.join(video_dir, $line)")

  parser.add_argument("--out_dir", default=None,
                      help="out_dir/$basename/%%d.json, start from 0 index. "
                           "This is the object box output. Leave this blank "
                           "when use tracking to avoid saving the obj class "
                           "output to save IO time.")

  parser.add_argument("--frame_gap", default=8, type=int)

  parser.add_argument("--threshold_conf", default=0.0001, type=float)

  parser.add_argument("--is_load_from_pb", action="store_true",
                      help="load from a frozen graph")
  parser.add_argument("--log_time_and_gpu", action="store_true")

  # ------ for box feature extraction
  parser.add_argument("--get_box_feat", action="store_true",
                      help="this will generate (num_box, 256, 7, 7) tensor for "
                           "each frame")
  parser.add_argument("--box_feat_path", default=None,
                      help="output will be out_dir/$basename/%%d.npy, start "
                      "from 0 index")

  parser.add_argument("--version", type=int, default=4, help="model version")
  parser.add_argument("--is_coco_model", action="store_true",
                      help="is coco model, will output coco classes instead")

  # ---- gpu params
  parser.add_argument("--gpu", default=1, type=int, help="number of gpu")
  parser.add_argument("--gpuid_start", default=0, type=int,
                      help="start of gpu id")
  parser.add_argument("--im_batch_size", type=int, default=1)
  parser.add_argument("--use_all_mem", action="store_true")

  # --- for internal visualization
  parser.add_argument("--visualize", action="store_true")
  parser.add_argument("--vis_path", default=None)
  parser.add_argument("--vis_thres", default=0.7, type=float)

  # ----------- model params
  parser.add_argument("--num_class", type=int, default=15,
                      help="num catagory + 1 background")

  parser.add_argument("--model_path", default="/app/object_detection_model")

  parser.add_argument("--rpn_batch_size", type=int, default=256,
                      help="num roi per image for RPN  training")
  parser.add_argument("--frcnn_batch_size", type=int, default=512,
                      help="num roi per image for fastRCNN training")

  parser.add_argument("--rpn_test_post_nms_topk", type=int, default=1000,
                      help="test post nms, input to fast rcnn")

  parser.add_argument("--max_size", type=int, default=1920,
                      help="num roi per image for RPN and fastRCNN training")
  parser.add_argument("--short_edge_size", type=int, default=1080,
                      help="num roi per image for RPN and fastRCNN training")

  # use lijun video loader, this should deal with avi videos
  # with duplicate frames
  parser.add_argument(
      "--use_lijun_video_loader", action="store_true",
      help="use video loader from https://github.com/Lijun-Yu/diva_io")
  parser.add_argument("--use_moviepy", action="store_true")


  # ----------- tracking params
  parser.add_argument("--get_tracking", action="store_true",
                      help="this will generate tracking results for each frame")
  parser.add_argument("--tracking_dir", default="/tmp",
                      help="output will be out_dir/$videoname.txt, start from 0"
                           " index")
  parser.add_argument("--tracking_objs", default="Person,Vehicle",
                      help="Objects to be tracked, default are Person and "
                           "Vehicle")
  parser.add_argument("--min_confidence", default=0.85, type=float,
                      help="Detection confidence threshold. Disregard all "
                           "detections that have a confidence lower than this "
                           "value.")
  parser.add_argument("--min_detection_height", default=0, type=int,
                      help="Threshold on the detection bounding box height. "
                           "Detections with height smaller than this value are "
                           "disregarded")
  # this does not make a big difference
  parser.add_argument("--nms_max_overlap", default=0.85, type=float,
                      help="Non-maxima suppression threshold: Maximum detection"
                           " overlap.")

  parser.add_argument("--max_iou_distance", type=float, default=0.5,
                      help="Iou distance for tracker.")
  parser.add_argument("--max_cosine_distance", type=float, default=0.5,
                      help="Gating threshold for cosine distance metric (object"
                           " appearance).")
  # nn_budget smaller more tracks
  parser.add_argument("--nn_budget", type=int, default=5,
                      help="Maximum size of the appearance descriptors gallery."
                           " If None, no budget is enforced.")

  parser.add_argument("--bupt_exp", action="store_true",
                      help="activity box experiemnt")

  # ---- tempory: for activity detection model
  parser.add_argument("--actasobj", action="store_true")
  parser.add_argument("--actmodel_path",
                      default="/app/activity_detection_model")

  parser.add_argument("--resnet152", action="store_true", help="")
  parser.add_argument("--resnet50", action="store_true", help="")
  parser.add_argument("--resnet34", action="store_true", help="")
  parser.add_argument("--resnet18", action="store_true", help="")
  parser.add_argument("--use_se", action="store_true",
                      help="use squeeze and excitation in backbone")
  parser.add_argument("--use_frcnn_class_agnostic", action="store_true",
                      help="use class agnostic fc head")
  parser.add_argument("--use_resnext", action="store_true", help="")
  parser.add_argument("--use_att_frcnn_head", action="store_true",
                      help="use attention to sum [K, 7, 7, C] feature into"
                           " [K, C]")

  # ------ 04/2020, efficientdet
  parser.add_argument("--is_efficientdet", action="store_true")
  parser.add_argument("--efficientdet_modelname", default="efficientdet-d0")
  parser.add_argument("--efficientdet_max_detection_topk", type=int,
                      default=5000, help="#topk boxes before NMS")
  parser.add_argument("--efficientdet_min_level", type=int, default=3)
  parser.add_argument("--efficientdet_max_level", type=int, default=7)

  # ---- COCO Mask-RCNN model
  parser.add_argument("--add_mask", action="store_true")

  # --------------- exp junk
  parser.add_argument("--use_dilations", action="store_true",
                      help="use dilations=2 in res5")
  parser.add_argument("--use_deformable", action="store_true",
                      help="use deformable conv")
  parser.add_argument("--add_act", action="store_true",
                      help="add activitiy model")
  parser.add_argument("--finer_resolution", action="store_true",
                      help="fpn use finer resolution conv")
  parser.add_argument("--fix_fpn_model", action="store_true",
                      help="for finetuneing a fpn model, whether to fix the"
                           " lateral and poshoc weights")
  parser.add_argument("--is_cascade_rcnn", action="store_true",
                      help="cascade rcnn on top of fpn")
  parser.add_argument("--add_relation_nn", action="store_true",
                      help="add relation network feature")

  parser.add_argument("--test_frame_extraction", action="store_true")
  parser.add_argument("--use_my_naming", action="store_true")

  # for efficient use of COCO model classes
  parser.add_argument("--use_partial_classes", action="store_true")

  args = parser.parse_args()

  if args.use_partial_classes:
    assert args.is_coco_model
    args.partial_classes = [classname for classname in coco_obj_to_actev_obj]

  assert args.gpu == args.im_batch_size  # one gpu one image
  assert args.gpu == 1, "Currently only support single-gpu inference"

  if args.is_load_from_pb:
    args.load_from = args.model_path

  args.controller = "/cpu:0"  # parameter server

  targetid2class = targetid2class
  targetClass2id = targetClass2id

  if args.actasobj:
    from class_ids import targetAct2id
    targetClass2id = targetAct2id
    targetid2class = {targetAct2id[one]: one for one in targetAct2id}
  if args.bupt_exp:
    from class_ids import targetAct2id_bupt
    targetClass2id = targetAct2id_bupt
    targetid2class = {targetAct2id_bupt[one]: one for one in targetAct2id_bupt}

  assert len(targetClass2id) == args.num_class, (len(targetClass2id),
                                                 args.num_class)


  assert args.version in [2, 3, 4, 5, 6], \
         "Currently we only have version 2-6 model"

  if args.version == 2:
    pass
  elif args.version == 3:
    args.use_dilations = True
  elif args.version == 4:
    args.use_frcnn_class_agnostic = True
    args.use_dilations = True
  elif args.version == 5:
    args.use_frcnn_class_agnostic = True
    args.use_dilations = True
  elif args.version == 6:
    args.use_frcnn_class_agnostic = True
    args.use_se = True

  if args.is_coco_model:
    assert args.version == 2
    targetClass2id = coco_obj_class_to_id
    targetid2class = coco_obj_id_to_class
    args.num_class = 81
    if args.use_partial_classes:
      partial_classes = ["BG"] + args.partial_classes
      targetClass2id = {classname: i
                        for i, classname in enumerate(partial_classes)}
      targetid2class = {targetClass2id[o]: o for o in targetClass2id}

  # ---- 04/2020, efficientdet
  if args.is_efficientdet:
    targetClass2id = coco_obj_class_to_id
    targetid2class = coco_obj_id_to_class
    args.num_class = 81
    args.is_coco_model = True

  # ---------------more defautls
  args.is_pack_model = False
  args.diva_class3 = True
  args.diva_class = False
  args.diva_class2 = False
  args.use_small_object_head = False
  args.use_so_score_thres = False
  args.use_so_association = False
  args.use_gn = False
  args.so_person_topk = 10
  args.use_conv_frcnn_head = False
  args.use_cpu_nms = False
  args.use_bg_score = False
  args.freeze_rpn = True
  args.freeze_fastrcnn = True
  args.freeze = 2
  args.small_objects = ["Prop", "Push_Pulled_Object",
                        "Prop_plus_Push_Pulled_Object", "Bike"]
  args.no_obj_detect = False
  #args.add_mask = False
  args.is_fpn = True
  # args.new_tensorpack_model = True
  args.mrcnn_head_dim = 256
  args.is_train = False

  args.rpn_min_size = 0
  args.rpn_proposal_nms_thres = 0.7
  args.anchor_strides = (4, 8, 16, 32, 64)

  # [3] is 32, since we build FPN with r2,3,4,5, so 2**5
  args.fpn_resolution_requirement = float(args.anchor_strides[3])

  if args.is_efficientdet:
    args.fpn_resolution_requirement = 128.0  # 2 ** max_level
    args.short_edge_size = np.ceil(
        args.short_edge_size / args.fpn_resolution_requirement) * \
            args.fpn_resolution_requirement
  args.max_size = np.ceil(args.max_size / args.fpn_resolution_requirement) * \
                  args.fpn_resolution_requirement

  args.fpn_num_channel = 256

  args.fpn_frcnn_fc_head_dim = 1024

  # ---- all the mask rcnn config

  args.resnet_num_block = [3, 4, 23, 3]  # resnet 101
  args.use_basic_block = False  # for resnet-34 and resnet-18
  if args.resnet152:
    args.resnet_num_block = [3, 8, 36, 3]
  if args.resnet50:
    args.resnet_num_block = [3, 4, 6, 3]
  if args.resnet34:
    args.resnet_num_block = [3, 4, 6, 3]
    args.use_basic_block = True
  if args.resnet18:
    args.resnet_num_block = [2, 2, 2, 2]
    args.use_basic_block = True

  args.anchor_stride = 16  # has to be 16 to match the image feature
  args.anchor_sizes = (32, 64, 128, 256, 512)

  args.anchor_ratios = (0.5, 1, 2)

  args.num_anchors = len(args.anchor_sizes) * len(args.anchor_ratios)
  # iou thres to determine anchor label
  # args.positive_anchor_thres = 0.7
  # args.negative_anchor_thres = 0.3

  # when getting region proposal, avoid getting too large boxes
  args.bbox_decode_clip = np.log(args.max_size / 16.0)

  # fastrcnn
  args.fastrcnn_batch_per_im = args.frcnn_batch_size
  args.fastrcnn_bbox_reg_weights = np.array([10, 10, 5, 5], dtype="float32")

  args.fastrcnn_fg_thres = 0.5  # iou thres
  # args.fastrcnn_fg_ratio = 0.25 # 1:3 -> pos:neg

  # testing
  args.rpn_test_pre_nms_topk = 6000

  args.fastrcnn_nms_iou_thres = 0.5

  args.result_score_thres = args.threshold_conf
  args.result_per_im = 100

  return args


def initialize(config, sess):
  """
    load the tf model weights into session
  """
  tf.global_variables_initializer().run()
  allvars = tf.global_variables()
  allvars = [var for var in allvars if "global_step" not in var.name]
  restore_vars = allvars
  opts = ["Adam", "beta1_power", "beta2_power", "Adam_1", "Adadelta_1",
          "Adadelta", "Momentum"]
  restore_vars = [var for var in restore_vars
                  if var.name.split(":")[0].split("/")[-1] not in opts]

  saver = tf.train.Saver(restore_vars, max_to_keep=5)

  load_from = config.model_path
  ckpt = tf.train.get_checkpoint_state(load_from)
  if ckpt and ckpt.model_checkpoint_path:
    loadpath = ckpt.model_checkpoint_path
    saver.restore(sess, loadpath)
  else:
    if os.path.exists(load_from):
      if load_from.endswith(".ckpt"):
        # load_from should be a single .ckpt file
        saver.restore(sess, load_from)
      elif load_from.endswith(".npz"):
        # load from dict
        weights = np.load(load_from)
        params = {get_op_tensor_name(n)[1]:v
                  for n, v in dict(weights).iteritems()}
        param_names = set(params.iterkeys())

        variables = restore_vars

        variable_names = set([k.name for k in variables])

        intersect = variable_names & param_names

        restore_vars = [v for v in variables if v.name in intersect]

        with sess.as_default():
          for v in restore_vars:
            vname = v.name
            v.load(params[vname])

        not_used = [(o, weights[o].shape)
                    for o in weights.keys()
                    if get_op_tensor_name(o)[1] not in intersect]
        if not not_used:
          print("warning, %s/%s in npz not restored:%s" % (
              len(weights.keys()) - len(intersect), len(weights.keys()),
              not_used))

      else:
        raise Exception("Not recognized model type:%s" % load_from)
    else:
      raise Exception("Model not exists")


def check_args(args):
  """Check the argument."""
  assert args.video_dir is not None
  assert args.video_lst_file is not None
  assert args.frame_gap >= 1
  if args.get_box_feat:
    assert args.box_feat_path is not None
    if not os.path.exists(args.box_feat_path):
      os.makedirs(args.box_feat_path)
  #print("cv2 version %s" % (cv2.__version__)


gpu_util_logs = []
gpu_temp_logs = []


def log_gpu_util(interval, gpuid_range):
  """
    A function to keep track of gpu usage using nvidia-smi
  """
  global gpu_util_logs
  while True:
    time.sleep(interval)
    gpu_temps, gpu_utils = parse_nvidia_smi(gpuid_range)
    gpu_util_logs.extend(gpu_utils)
    gpu_temp_logs.extend(gpu_temps)


if __name__ == "__main__":
  args = get_args()

  if args.log_time_and_gpu:
    gpu_log_interval = 10 # every k seconds
    start_time = time.time()
    gpu_check_thread = threading.Thread(
        target=log_gpu_util,
        args=[gpu_log_interval, (args.gpuid_start, args.gpu)])
    gpu_check_thread.daemon = True
    gpu_check_thread.start()

  check_args(args)

  videolst = [os.path.join(args.video_dir, one.strip())
              for one in open(args.video_lst_file).readlines()]

  if args.out_dir is not None:
    if not os.path.exists(args.out_dir):
      os.makedirs(args.out_dir)

  # 2020, deal with opencv  avi video "bug":
  # https://github.com/opencv/opencv/issues/9053
  # need pyav
  if args.use_lijun_video_loader:
    # https://github.com/Lijun-Yu/diva_io
    from diva_io.video import VideoReader

  if args.use_moviepy:
    from moviepy.editor import VideoFileClip

  if args.visualize:
    from viz import draw_boxes

    vis_path = args.vis_path
    if not os.path.exists(vis_path):
      os.makedirs(vis_path)

  # 1. load the object detection model
  model = get_model(args, args.gpuid_start, controller=args.controller)

  tfconfig = tf.ConfigProto(allow_soft_placement=True)
  if not args.use_all_mem:
    tfconfig.gpu_options.allow_growth = True
  tfconfig.gpu_options.visible_device_list = "%s" % (
      ",".join(["%s" % i
                for i in range(args.gpuid_start, args.gpuid_start + args.gpu)]))

  with tf.Session(config=tfconfig) as sess:

    if not args.is_load_from_pb:
      initialize(config=args, sess=sess)

    for videofile in tqdm(videolst, ascii=True):
      # 2. read the video file
      if args.use_lijun_video_loader:
        vcap = VideoReader(videofile)
        frame_count = int(vcap.length)
      elif args.use_moviepy:
        vcap = VideoFileClip(videofile, audio=False)
        frame_count = int(vcap.fps * vcap.duration)  # uh
        vcap_iter = vcap.iter_frames()
      else:
        try:
          vcap = cv2.VideoCapture(videofile)
          if not vcap.isOpened():
            raise Exception("cannot open %s" % videofile)
        except Exception as e:
          # raise e
          # just move on to the next video
          print("warning, cannot open %s" % videofile)
          continue

        # opencv 2
        if cv2.__version__.split(".")[0] == "2":
          frame_count = vcap.get(cv2.cv.CV_CAP_PROP_FRAME_COUNT)
        else:
          # opencv 3/4
          frame_count = vcap.get(cv2.CAP_PROP_FRAME_COUNT)

      # initialize tracking module
      if args.get_tracking:
        tracking_objs = args.tracking_objs.split(",")
        tracker_dict = {}
        tracking_results_dict = {}
        tmp_tracking_results_dict = {}
        for tracking_obj in tracking_objs:
          metric = metric = nn_matching.NearestNeighborDistanceMetric(
              "cosine", args.max_cosine_distance, args.nn_budget)
          tracker_dict[tracking_obj] = Tracker(
              metric, max_iou_distance=args.max_iou_distance)
          tracking_results_dict[tracking_obj] = []
          tmp_tracking_results_dict[tracking_obj] = {}

      # videoname = os.path.splitext(os.path.basename(videofile))[0]
      videoname = os.path.basename(videofile)
      if args.out_dir is not None:  # not saving box json to save time
        video_out_path = os.path.join(args.out_dir, videoname)
        if not os.path.exists(video_out_path):
          os.makedirs(video_out_path)

      # for box feature
      if args.get_box_feat:
        feat_out_path = os.path.join(args.box_feat_path, videoname)
        if not os.path.exists(feat_out_path):
          os.makedirs(feat_out_path)

      # 3. read frame one by one
      cur_frame = 0
      vis_count = 0
      frame_stack = []
      while cur_frame < frame_count:
        if args.use_moviepy:
          suc = True
          frame = next(vcap_iter)
        else:
          suc, frame = vcap.read()
        if not suc:
          cur_frame += 1
          tqdm.write("warning, %s frame of %s failed" % (cur_frame, videoname))
          continue

        # skip some frame if frame_gap >1
        if cur_frame % args.frame_gap != 0:
          cur_frame += 1
          continue

        # 4. run detection on the frame stack if there is enough
        if args.use_moviepy:
          # moviepy ask ffmpeg to get rgb24
          frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
        im = frame.astype("float32")

        if args.test_frame_extraction:
          frame_file = os.path.join(video_out_path, "%d.jpg" % cur_frame)
          cv2.imwrite(frame_file, im)
          cur_frame += 1
          continue

        resized_image = resizeImage(im, args.short_edge_size, args.max_size)

        scale = (resized_image.shape[0] * 1.0 / im.shape[0] + \
                 resized_image.shape[1] * 1.0 / im.shape[1]) / 2.0

        feed_dict = model.get_feed_dict_forward(resized_image)

        if args.get_box_feat:
          sess_input = [model.final_boxes, model.final_labels,
                        model.final_probs, model.fpn_box_feat]

          final_boxes, final_labels, final_probs, box_feats = sess.run(
              sess_input, feed_dict=feed_dict)
          assert len(box_feats) == len(final_boxes)
          # save the box feature first

          featfile = os.path.join(feat_out_path, "%d.npy" % (cur_frame))
          np.save(featfile, box_feats)
        elif args.get_tracking:

          if args.add_mask:
            sess_input = [model.final_boxes, model.final_labels,
                          model.final_probs, model.fpn_box_feat,
                          model.final_masks]
            final_boxes, final_labels, final_probs, box_feats, final_masks = \
                sess.run(sess_input, feed_dict=feed_dict)
          else:
            sess_input = [model.final_boxes, model.final_labels,
                          model.final_probs, model.fpn_box_feat]
            final_boxes, final_labels, final_probs, box_feats = sess.run(
                sess_input, feed_dict=feed_dict)
            if args.is_efficientdet:
              # the output here is 1 - num_partial_classes
              if args.use_partial_classes:
                for i in range(len(final_labels)):
                  final_labels[i] = coco_obj_class_to_id[
                      args.partial_classes[final_labels[i] - 1]]
              else:
                # 1-90 to 1-80
                for i in range(len(final_labels)):
                  final_labels[i] = \
                      coco_obj_class_to_id[coco_id_mapping[final_labels[i]]]

          assert len(box_feats) == len(final_boxes)

          for tracking_obj in tracking_objs:
            target_tracking_obs = [tracking_obj]
            detections = create_obj_infos(
                cur_frame, final_boxes, final_probs, final_labels, box_feats,
                targetid2class, target_tracking_obs, args.min_confidence,
                args.min_detection_height, scale,
                is_coco_model=args.is_coco_model,
                coco_to_actev_mapping=coco_obj_to_actev_obj)
            # Run non-maxima suppression.
            boxes = np.array([d.tlwh for d in detections])
            scores = np.array([d.confidence for d in detections])
            indices = preprocessing.non_max_suppression(
                boxes, args.nms_max_overlap, scores)
            detections = [detections[i] for i in indices]

            # tracking
            tracker_dict[tracking_obj].predict()
            tracker_dict[tracking_obj].update(detections)

            # Store results
            for track in tracker_dict[tracking_obj].tracks:
              if not track.is_confirmed() or track.time_since_update > 1:
                if (not track.is_confirmed()) and track.time_since_update == 0:
                  bbox = track.to_tlwh()
                  if track.track_id not in \
                      tmp_tracking_results_dict[tracking_obj]:
                    tmp_tracking_results_dict[tracking_obj][track.track_id] = \
                        [[cur_frame, track.track_id, bbox[0], bbox[1],
                          bbox[2], bbox[3]]]
                  else:
                    tmp_tracking_results_dict[
                        tracking_obj][track.track_id].append(
                            [cur_frame, track.track_id,
                             bbox[0], bbox[1], bbox[2], bbox[3]])
                continue
              bbox = track.to_tlwh()
              if track.track_id in tmp_tracking_results_dict[tracking_obj]:
                pred_list = tmp_tracking_results_dict[tracking_obj][
                    track.track_id]
                for pred_data in pred_list:
                  tracking_results_dict[tracking_obj].append(pred_data)
                tmp_tracking_results_dict[tracking_obj].pop(track.track_id,
                                                            None)
              tracking_results_dict[tracking_obj].append([
                  cur_frame, track.track_id, bbox[0], bbox[1], bbox[2],
                  bbox[3]])

        else:
          if args.add_mask:
            sess_input = [model.final_boxes, model.final_labels,
                          model.final_probs, model.final_masks]
            final_boxes, final_labels, final_probs, final_masks = sess.run(
                sess_input, feed_dict=feed_dict)
          else:
            sess_input = [model.final_boxes, model.final_labels,
                          model.final_probs]
            final_boxes, final_labels, final_probs = sess.run(
                sess_input, feed_dict=feed_dict)

        if args.out_dir is None:
          cur_frame += 1
          continue

        # ---------------- get the json outputs for object detection

        # scale back the box to original image size
        final_boxes = final_boxes / scale

        if args.add_mask:
          final_masks = [fill_full_mask(box, mask, im.shape[:2])
                         for box, mask in zip(final_boxes, final_masks)]

        # save as json
        pred = []

        for j, (box, prob, label) in enumerate(zip(
            final_boxes, final_probs, final_labels)):
          box[2] -= box[0]
          box[3] -= box[1]  # produce x,y,w,h output

          cat_id = int(label)
          cat_name = targetid2class[cat_id]

          # encode mask
          rle = None
          if args.add_mask:
            final_mask = final_masks[j] # [14, 14]
            rle = cocomask.encode(np.array(
                final_mask[:, :, None], order="F"))[0]
            rle["counts"] = rle["counts"].decode("ascii")

          res = {
              "category_id": cat_id,
              "cat_name": cat_name,  # [0-80]
              "score": float(round(prob, 7)),
              #"bbox": list(map(lambda x: float(round(x, 2)), box)),
              "bbox": [float(round(x, 2)) for x in box],
              "segmentation": rle,
          }

          pred.append(res)

        if args.use_my_naming:
          predfile = os.path.join(
              video_out_path,
              "%s_F_%08d.json" % (os.path.splitext(videoname)[0], cur_frame))
        else:
          predfile = os.path.join(video_out_path, "%d.json" % (cur_frame))

        with open(predfile, "w") as f:
          json.dump(pred, f)

        # for visualization
        if args.visualize:
          good_ids = [i for i in range(len(final_boxes))
                      if final_probs[i] >= args.vis_thres]
          final_boxes, final_labels, final_probs = final_boxes[good_ids], \
              final_labels[good_ids], final_probs[good_ids]
          vis_boxes = np.asarray(
              [[box[0], box[1], box[2] + box[0], box[3] + box[1]]
               for box in final_boxes])
          vis_labels = ["%s_%.2f" % (targetid2class[cat_id], prob)
                        for cat_id, prob in zip(final_labels, final_probs)]
          newim = draw_boxes(im, vis_boxes, vis_labels,
                             color=np.array([255, 0, 0]), font_scale=0.5,
                             thickness=2)

          vis_file = os.path.join(vis_path,
                                  "%s_F_%08d.jpg" % (videoname, vis_count))
          cv2.imwrite(vis_file, newim)

        cur_frame += 1

      if not args.use_lijun_video_loader and not args.use_moviepy:
        vcap.release()

      if args.get_tracking:
        for tracking_obj in tracking_objs:
          output_dir = os.path.join(args.tracking_dir, videoname, tracking_obj)
          if not os.path.exists(output_dir):
            os.makedirs(output_dir)
          output_file = os.path.join(
              output_dir, "%s.txt" % (os.path.splitext(videoname))[0])

          tracking_results = sorted(tracking_results_dict[tracking_obj],
                                    key=lambda x: (x[0], x[1]))
          # print(len(tracking_results)
          tracking_data = np.asarray(tracking_results)
          # print(tracking_data.shape
          tracking_data = linear_inter_bbox(tracking_data, args.frame_gap)
          tracking_data = filter_short_objs(tracking_data)
          tracking_results = tracking_data.tolist()
          with open(output_file, "w") as fw:
            for row in tracking_results:
              line = "%d,%d,%.2f,%.2f,%.2f,%.2f,1,-1,-1,-1" % (
                  row[0], row[1], row[2], row[3], row[4], row[5])
              fw.write(line + "\n")

      if args.test_frame_extraction:
        tqdm.write(
            "video %s got %s frames, opencv said frame count is %s" % (
                videoname, cur_frame, frame_count))
  if args.log_time_and_gpu:
    end_time = time.time()
    print("total run time %s (%s), log gpu utilize every %s seconds and get "
          "median %.2f%% and average %.2f%%. GPU temperature median %.2f and "
          "average %.2f (C)" % (
              sec2time(end_time - start_time),
              end_time - start_time,
              gpu_log_interval,
              np.median(gpu_util_logs)*100,
              np.mean(gpu_util_logs)*100,
              np.median(gpu_temp_logs),
              np.mean(gpu_temp_logs)))
  cv2.destroyAllWindows()