def task1(gt_file):
gt = read_xml_gt(gt_file)
classes_to_keep = ['car']
gt = filter_gt(gt, classes_to_keep)
det_bb = generate_noisy_annotations(gt)
lst_gt = [item[0] for item in gt]
last_frame = np.max(lst_gt)
miou = 0
for f_val in range(0, last_frame):
frame_gt_bb = [gt[i] for i, num in enumerate(gt) if num[0] == f_val]
frame_det_bb = [det_bb[i] for i, num in enumerate(det_bb) if num[0] == f_val]
miou += frame_miou(frame_det_bb, frame_gt_bb, confidence=False)
miou = miou/last_frame
print("Noisy GT (Sorted Random): ", calculate_ap(det_bb, gt, 'random'))
print("Noisy GT (Sorted by Area): ", calculate_ap(det_bb, gt, 'area'))
preds_mask = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_mask_rcnn.txt")
print("maskrcnn ap: {}".format(calculate_ap(preds_mask, gt, 'sort')))
preds_ssd = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_ssd512.txt")
print("ssd ap: {}".format(calculate_ap(preds_ssd, gt, 'sort')))
preds_yolo = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_yolo3.txt")
print("yolo ap: {}".format(calculate_ap(preds_yolo, gt, 'sort')))
def task_31():
gt_annot_file = 'datasets/ai_challenge_s03_c010-full_annotation.xml'
frames_path = 'datasets/AICity_data/train/S03/c010/data'
det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/our_results_finetune_faster.txt")
gt_bb = read_xml_gt_options(gt_annot_file, False, False)
gt_bb = filter_gt(gt_bb, ["car"])
video_n_frames = number_of_images_jpg(frames_path)
det_bb_max_iou, idd = tracking_iou(frames_path, copy.deepcopy(det_bb), video_n_frames, mode='of')
ap_max_iou = calculate_ap(det_bb_max_iou, gt_bb, 0, video_n_frames, mode='sort')
print("Ap after tracking with maximum IoU: {}".format(ap_max_iou))
det_bb_max_iou, idd = tracking_iou(frames_path, copy.deepcopy(det_bb), video_n_frames, mode='other')
ap_max_iou = calculate_ap(det_bb_max_iou, copy.deepcopy(gt_bb), 0, video_n_frames, mode='sort')
print("Ap after tracking with maximum IoU: {}".format(ap_max_iou))
new_tracks = restore_tracks(frames_path, det_bb_max_iou)
ap_max_iou = calculate_ap(new_tracks, copy.deepcopy(gt_bb), 0, video_n_frames, mode='sort')
print("Ap after tracking with maximum IoU and of: {}".format(ap_max_iou))
ini_frame = 700
end_frame = 900
animation_tracks(new_tracks, idd, ini_frame, end_frame, frames_path)
def task22(gt_annot_file, detections_file, frames_path):
ap_mode = 'area'
#Read and filter detections
det_bb = read_detections_file(detections_file)
det_bb = filter_det_confidence(det_bb, threshold=0.5)
#Read and filter gt
gt_bb = read_xml_gt_options(gt_annot_file, False, False)
gt_bb = filter_gt(gt_bb, ["car"])
#Calculate original ap
video_n_frames = number_of_images_jpg(frames_path)
# Constant velocity
det_bb_vel = kalman_filter_tracking(det_bb, video_n_frames, 0)
# Constant acceleration
det_bb_acc = kalman_filter_tracking(det_bb, video_n_frames, 1)
#Print ap results
print("Original ap: {}".format(
calculate_ap(det_bb, gt_bb, 0, video_n_frames, mode=ap_mode)))
print("Ap after tracking with constant velocity: {}".format(
calculate_ap(det_bb_vel, gt_bb, 0, video_n_frames, mode=ap_mode)))
print("Ap after tracking with constant acceleration: {}".format(
calculate_ap(det_bb_acc, gt_bb, 0, video_n_frames, mode=ap_mode)))
#Obtain animation
max_id_vel = max(det_bb_vel, key=lambda x: x[2])[2]
max_id_acc = max(det_bb_acc, key=lambda x: x[2])[2]
ini_frame = 550
end_frame = 650
def task21(gt_annot_file, detections_file, frames_path):
ap_mode = 'area'
# Read and filter detections
# det_bb = read_detections_file(detections_file)
# det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_mask_rcnn.txt")
# det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_ssd512.txt")
# det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_yolo3.txt")
# det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/our_results_coco_faster.txt")
det_bb = read_detections_file(
"datasets/AICity_data/train/S03/c010/det/our_results_finetune_faster.txt"
)
det_bb = filter_det_confidence(det_bb, threshold=0.5)
#Read and filter gt
gt_bb = read_xml_gt_options(gt_annot_file, False, False)
gt_bb = filter_gt(gt_bb, ["car"])
video_n_frames = number_of_images_jpg(frames_path)
original_ap = calculate_ap(det_bb, gt_bb, 0, video_n_frames, mode=ap_mode)
det_bb_max_iou, idd = tracking_iou(copy.deepcopy(det_bb), video_n_frames)
ap_max_iou = calculate_ap(det_bb_max_iou,
gt_bb,
0,
video_n_frames,
mode=ap_mode)
print("Original ap: {}".format(original_ap))
print("Ap after tracking with maximum IoU: {}".format(ap_max_iou))
ini_frame = 550
end_frame = 650
animation_tracks(det_bb_max_iou, idd, ini_frame, end_frame, frames_path)
def task1(frames_path,
detections_file,
gt_file,
tracking_method,
postprocessing=True,
visualization=False):
print("Start loading data")
# Read gt
gt_bb = read_gt_txt(gt_file)
# Read and filter detections
det_bb = read_detections_file(detections_file)
det_bb = filter_det_confidence(det_bb, threshold=0.5)
# Tracking
print("Start tracking")
video_n_frames = number_of_images_jpg(frames_path)
if tracking_method == "MaxOverlap":
det_bb_tracking, id_max = tracking_iou(frames_path,
copy.deepcopy(det_bb),
video_n_frames,
mode='other')
elif tracking_method == "Kalman":
det_bb_tracking = kalman_filter_tracking(copy.deepcopy(det_bb),
video_n_frames, 0)
id_max = max([v[2] for v in det_bb_tracking])
else:
raise NotImplemented
# Postprocessing
if postprocessing:
print("Starting postprocessing tracking")
det_bb_tracking = clean_tracks(copy.deepcopy(det_bb_tracking), id_max)
det_bb_tracking = remove_parked(copy.deepcopy(det_bb_tracking),
id_max,
threshold=10.0)
# Results
print("Start obtaining metrics")
ap = calculate_ap(det_bb_tracking,
copy.deepcopy(gt_bb),
0,
video_n_frames,
mode='sort')
print("Average precision: {}".format(ap))
# Create animation if required
if visualization:
print("Start storing animation")
ini_frame = 700
end_frame = 800
frames_path = 'datasets/AICity_data/train/S03/c010/data'
animation_tracks(det_bb_tracking, id_max, ini_frame, end_frame,
frames_path)
return det_bb_tracking
def task2(gt_file, ini_frame):
gt_bb = read_xml_gt(gt_file)
classes_to_keep = ['car']
gt_bb = filter_gt(gt_bb, classes_to_keep)
# det_bb = generate_noisy_annotations(gt_bb)
# det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_mask_rcnn.txt")
# det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_ssd512.txt")
# det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_yolo3.txt")
det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/our_results_finetune_faster.txt")
fps = 10
seconds = 30
animation_2bb("bb_frames_our_faster", ".gif", gt_bb, det_bb, "datasets/AICity_data/train/S03/c010/data/", fps, seconds,
ini_frame, int(1920 / 4), int(1080 / 4))
mious = []
for f_val in range(ini_frame, ini_frame+int(seconds*fps)):
frame_gt_bb = [gt_bb[i] for i, num in enumerate(gt_bb) if num[0] == f_val]
frame_det_bb = [det_bb[i] for i, num in enumerate(det_bb) if num[0] == f_val]
mious.append(frame_miou(frame_det_bb, frame_gt_bb, confidence=False))
xx = np.arange(ini_frame, ini_frame+int(seconds*fps))
plt.figure()
plt.ylim((0,1))
plt.xlabel('Frame')
plt.ylabel('mIoU')
## plt.plot(xx, yolo, label = 'YOLOv3')
## plt.plot(xx, mask, label = 'Mask R-CNN')
## plt.plot(xx, ssd, label = 'SSD')
plt.plot(xx, retina, label = 'RetinaNet')
plt.plot(xx, faster, label = 'Faster R-CNN')
plt.plot(xx, our_faster, label = 'Faster R-CNN (Fine tuned)')
plt.legend()
frames = list(range(0, len(mious)))
ani = plot_animation(frames, mious, 'Frame', 'IoU', [0,1], fps)
ani.save('iou_our_faster.gif')