def evaluate(self, eval_loader):
mAP, aps, pr_curves = eval_fcos_det(self.cfg,
self.criterion,
eval_loader,
self.model,
rescale=None)
logging.info(f"score_threshold:{self.cfg.TEST.SCORE_THRESHOLD}, "
f"nms_iou:{self.cfg.TEST.NMS_THRESHOLD}")
logging.info(f'mean Average Precision Across All Classes:{mAP}')
if mAP > self.best_performance or True:
# save PR curve
if self.cfg.TEST.PR_CURVE:
for class_idx, pr_curve in enumerate(pr_curves):
self.writer.add_figure(f'PR_curve/class_{class_idx + 1}',
pr_curve, self.iter)
logging.info('PR curves saved')
# save detected result
if self.cfg.TEST.IMAGE_DIR != '':
transform = PredictionTransform(resize=cfg.TEST.TEST_SIZE)
run_fcos_det_example(self.cfg, self.criterion,
self.cfg.TEST.IMAGE_DIR, transform,
self.model)
logging.info('test detected result saved')
return mAP
8: 'Misc',
}
for label in range(1, cfg.DATASET.NUM_CLASSES):
det_result[label] = {}
crop_im[label] = {}
for i, jpg_path in tqdm(enumerate(jpg_paths)):
if jpg_path.split('.')[-1].lower() not in [
'jpg', 'png', 'jpeg', 'bmp'
]:
continue
frame_id = int(jpg_path.split('.')[-2].split('/')[-1]) # start from 0
# img:(min_size, max_size, 3), boxes:(obj_num, 4), labels:(obj_num,) numpy type<list> list of array(1)
det_output = run_fcos_det_example(cfg,
criterion,
jpg_path,
det_transform,
detector,
ap_transform=ap_transform,
crop_transform=crop_transform)
if det_output is None:
continue
img, crop_img, boxes, labels = det_output
boxes[:, 2] = boxes[:, 2] - boxes[:, 0] # w
boxes[:, 3] = boxes[:, 3] - boxes[:, 1] # h
img = img.cuda(non_blocking=True)
img_embs = get_embeddings(emb, img).cpu().data.numpy()
assert len(img_embs) == len(crop_img) == len(boxes) == len(labels)
for label in range(1, cfg.DATASET.NUM_CLASSES):
label_idx = sorted(list(np.where(labels == label)[0]))
obj_num = len(label_idx)
# {'frame_id': (obj_num, crop_min, crop_max, 3)}
os.mkdir(video_crop_path)
video_path = osp.join(args.videos_root, video_name)
jpg_paths = []
# read test images
if osp.isfile(video_path):
jpg_paths = [video_path]
elif osp.isdir(video_path):
file_names = os.listdir(video_path)
jpg_paths = [
osp.join(video_path, file_name) for file_name in file_names
]
else:
raise FileNotFoundError
for i, jpg_path in tqdm(enumerate(jpg_paths)):
if jpg_path.split('.')[-1].lower() not in [
'jpg', 'png', 'jpeg', 'bmp'
]:
continue
frame_id = int(jpg_path.split('/')[-1].split('.')[-2])
frame_crop_path = osp.join(video_crop_path, str(frame_id))
if not osp.exists(frame_crop_path):
os.mkdir(frame_crop_path)
img, boxes, labels = run_fcos_det_example(cfg,
criterion,
jpg_path,
transform,
model,
is_crop=True)
write_crop(img, boxes, labels, frame_crop_path)
print(f"==> checkpoint do not exists: \"{cfg.MODEL.RESUME_PATH}\"")
raise FileNotFoundError
if not osp.exists(cfg.TEST.OUT_DIR):
os.mkdir(cfg.TEST.OUT_DIR)
# read test images
if osp.isfile(args.jpg_path):
jpg_paths = [args.jpg_path]
elif osp.isdir(args.jpg_path):
file_names = os.listdir(args.jpg_path)
jpg_paths = [
osp.join(args.jpg_path, file_name) for file_name in file_names
]
else:
raise FileNotFoundError
transform = PredictionTransform(size=(352, 1216))
timer = []
for i, jpg_path in tqdm(enumerate(jpg_paths)):
if jpg_path.split('.')[-1].lower() not in [
'jpg', 'png', 'jpeg', 'bmp'
]:
continue
start = time.time()
img = run_fcos_det_example(cfg, criterion, jpg_path, transform, model)
timer.append(time.time() - start)
num = jpg_path.split('.')[-2].split('/')[-1]
cv2.imwrite(osp.join(cfg.TEST.OUT_DIR, f'run_fcos_example_{num}.jpg'),
img)
print(f'finnal output dir is {cfg.TEST.OUT_DIR}')
print(f'average inference time: {sum(timer) / len(timer)}s')
det_transform = PredictionTransform(size=cfg.TEST.TEST_SIZE)
ap_transform = FacenetInferenceTransform(size=(cfg.TEST.CROP_SZIE))
# get det result dict using the detector
det_result = {}
for i, jpg_path in tqdm(enumerate(jpg_paths)):
if jpg_path.split('.')[-1].lower() not in [
'jpg', 'png', 'jpeg', 'bmp'
]:
continue
frame_id = int(jpg_path.split('.')[-2].split('/')[-1]) # start from 0
# img:(min_size, max_size, 3), boxes:(obj_num, 4), labels:(obj_num,) numpy type, mean_color <list> list of array(3)
img, boxes, labels, mean_color = run_fcos_det_example(
cfg,
criterion,
jpg_path,
det_transform,
detector,
ap_transform=ap_transform,
)
boxes[:, 2] = boxes[:, 2] - boxes[:, 0] # w
boxes[:, 3] = boxes[:, 3] - boxes[:, 1] # h
img = img.cuda(non_blocking=True)
img_embs = get_embeddings(emb, img).cpu().data.numpy()
assert len(img_embs) == len(boxes) == len(labels) == len(mean_color)
obj_num = len(img_embs)
# {'frame_id': (obj_num, emb_size+4+1+3) emb x y w h label(float) mean_color(array(3))}
det_result[frame_id] = np.zeros((obj_num, emb_size + 4 + 1 + 3))
det_result[frame_id][:, :emb_size] = img_embs
det_result[frame_id][:, emb_size:emb_size + 4] = boxes
det_result[frame_id][:, emb_size + 4] = labels
det_result[frame_id][:, emb_size + 5:] = mean_color