def _eval(self):
all_results_1st = eval_coco(self.df, lambda img: detect_one_image(img, self.pred_1st))
all_results_2nd = eval_coco(self.df, lambda img: detect_one_image(img, self.pred_2nd))
all_results_3rd = eval_coco(self.df, lambda img: detect_one_image(img, self.pred_3rd))
output_file_1st = os.path.join(
logger.get_logger_dir(), '1st_outputs{}.json'.format(self.global_step))
output_file_2nd = os.path.join(
logger.get_logger_dir(), '2nd_outputs{}.json'.format(self.global_step))
output_file_3rd = os.path.join(
logger.get_logger_dir(), '3rd_outputs{}.json'.format(self.global_step))
with open(output_file_1st, 'w') as f:
json.dump(all_results_1st, f)
with open(output_file_2nd, 'w') as f:
json.dump(all_results_2nd, f)
with open(output_file_3rd, 'w') as f:
json.dump(all_results_3rd, f)
try:
scores_1st = print_evaluation_scores(output_file_1st)
scores_2nd = print_evaluation_scores(output_file_2nd)
scores_3rd = print_evaluation_scores(output_file_3rd)
except Exception:
logger.exception("Exception in COCO evaluation.")
scores = {}
for k, v in scores_1st.items():
self.trainer.monitors.put_scalar(k, v)
for k, v in scores_2nd.items():
self.trainer.monitors.put_scalar(k, v)
for k, v in scores_3rd.items():
self.trainer.monitors.put_scalar(k, v)
def _eval(self):
all_results = eval_on_dataflow(
self.df, lambda img: detect_one_image(img, self.pred))
output_file = os.path.join(logger.get_logger_dir(),
'outputs{}.json'.format(self.global_step))
with open(output_file, 'w') as f:
json.dump(all_results, f)
def predict(pred_func, input_file):
img = cv2.imread(input_file, cv2.IMREAD_COLOR)
results = detect_one_image(img, pred_func)
print(len(results))
final = draw_final_outputs(img, results)
viz = np.concatenate((img, final), axis=1)
tpviz.interactive_imshow(viz)
def predict(pred_func, input_file):
img = cv2.imread(input_file, cv2.IMREAD_COLOR)
results = detect_one_image(img, pred_func)
final = draw_final_outputs(img, results)
viz = np.concatenate((img, final), axis=1)
# tpviz.interactive_imshow(viz)
cv2.imwrite(os.path.basename(input_file), viz)
def offline_evaluate(pred_func, output_file):
df = get_eval_dataflow()
all_results = eval_coco(
df, lambda img: detect_one_image(img, pred_func))
with open(output_file, 'w') as f:
json.dump(all_results, f)
print_evaluation_scores(output_file)
def _eval(self):
all_results, local_score = eval_on_dataflow(
self.df, lambda img: detect_one_image(img, self.pred))
scores = {}
scores['local'] = local_score
for k, v in scores.items():
self.trainer.monitors.put_scalar(k, v)
def offline_evaluate(pred_func, output_file):
df = get_eval_dataflow()
all_results = eval_on_dataflow(
df, lambda img: detect_one_image(img, pred_func))
with open(output_file, 'w') as f:
json.dump(all_results, f)
print_evaluation_scores(output_file)
def predict(pred_func, input_file):
img = cv2.imread(input_file, cv2.IMREAD_COLOR)
results = detect_one_image(img, pred_func)
final = draw_final_outputs(img, results)
viz = np.concatenate((img, final), axis=1)
# tpviz.interactive_imshow(viz)
cv2.imwrite('../drive/train_log/predicted_image.jpg', viz)
cv2.destroyAllWindows()
def predict(pred_func, input_file):
img = cv2.imread(input_file, cv2.IMREAD_COLOR)
results = detect_one_image(img, pred_func)
final = draw_final_outputs(img, results)
viz = np.concatenate((img, final), axis=1)
cv2.imwrite("output.png", viz)
logger.info("Inference output written to output.png")
tpviz.interactive_imshow(viz)
def _eval(self):
all_results = eval_coco(self.df, lambda img: detect_one_image(img, self.pred))
output_file = os.path.join(
logger.get_logger_dir(), 'outputs{}.json'.format(self.global_step))
with open(output_file, 'w') as f:
json.dump(all_results, f)
scores = print_evaluation_scores(output_file)
for k, v in scores.items():
self.trainer.monitors.put_scalar(k, v)
def predict_many(pred_func, input_files):
if not os.path.exists('output'):
os.mkdir('output')
for idx, input_file in enumerate(input_files):
img = cv2.imread(input_file, cv2.IMREAD_COLOR)
results = detect_one_image(img, pred_func)
final = draw_final_outputs(img, results)
plt.imshow(final)
plt.savefig(os.path.join('output', str(idx) + '.png'))
def _eval(self):
all_results = eval_coco(self.df, lambda img: detect_one_image(img, self.pred))
output_file = os.path.join(
logger.get_logger_dir(), 'outputs{}.json'.format(self.global_step))
with open(output_file, 'w') as f:
json.dump(all_results, f)
scores = print_evaluation_scores(output_file)
for k, v in scores.items():
self.trainer.monitors.put_scalar(k, v)
def offline_evaluate(pred_func, output_file):
df = get_test_dataflow(add_mask=False)
all_results = eval_on_dataflow(
df, lambda img: detect_one_image(img, pred_func))
#print(all_results)
# input()
with open(output_file, 'w') as f:
json.dump(all_results, f, cls=MyEncoder)
ret = print_evaluation_scores(output_file)
print(ret)
def predict_one_image(input_file, pred_func):
print('Predicting '+input_file)
path = '/home/bonniehu/projects/def-jjclark/bonniehu/FasterRCNN/pred_out/'
img = cv2.imread(input_file, cv2.IMREAD_COLOR)
results = detect_one_image(img, pred_func)
trimap_path = 'trimap/'
trimaps = trimap_outputs(img, results, trimap_path, input_file.split('/')[-1][:-4])
final = draw_final_outputs(img, results)
#viz = np.concatenate((img, final), axis=1)
#not support in ubuntu
#tpviz.interactive_imshow(viz)
misc.imsave(path+input_file.split('/')[-1][:-4]+'_mask.png', final)
def _eval(self):
all_results, local_score = eval_on_dataflow(
self.df, lambda img: detect_one_image(img, self.pred))
"""
output_file = os.path.join(
logger.get_logger_dir(), 'outputs{}.json'.format(self.global_step))
with open(output_file, 'w') as f:
json.dump(all_results, f, cls=MyEncoder)
scores = print_evaluation_scores(output_file)
"""
scores = {}
scores['local'] = local_score
for k, v in scores.items():
self.trainer.monitors.put_scalar(k, v)
def predict(pred_func, input_file):
img = cv2.imread(input_file, cv2.IMREAD_COLOR)
t = time.time()
results = detect_one_image(img, pred_func)
t_final = time.time() - t
height, width, channels = img.shape
print("Time: ", t_final, " for ", height * width / 1e6,
" Mpixels , i.e. Width=", width, " Height=", height, " Channels=",
channels)
print("Throughput: ", height * width / 1e6 / t_final, " Mpixels/Sec")
final = draw_final_outputs(img, results)
viz = np.concatenate((img, final), axis=1)
tpviz.interactive_imshow(viz)
def predict(model_path, input_file):
pred = OfflinePredictor(
PredictConfig(model=Model(),
session_init=get_model_loader(model_path),
input_names=['image'],
output_names=[
'fastrcnn_fg_probs',
'fastrcnn_fg_boxes',
]))
img = cv2.imread(input_file, cv2.IMREAD_COLOR)
results = detect_one_image(img, pred)
final = draw_final_outputs(img, results)
viz = np.concatenate((img, final), axis=1)
tpviz.interactive_imshow(viz)
def offline_evaluate(model_path, output_file):
pred = OfflinePredictor(
PredictConfig(model=Model(),
session_init=get_model_loader(model_path),
input_names=['image'],
output_names=[
'fastrcnn_fg_probs',
'fastrcnn_fg_boxes',
]))
df = get_eval_dataflow()
df = PrefetchDataZMQ(df, 1)
all_results = eval_on_dataflow(df, lambda img: detect_one_image(img, pred))
with open(output_file, 'w') as f:
json.dump(all_results, f)
print_evaluation_scores(output_file)
def offline_evaluate(pred_config, output_file):
num_gpu = cfg.TRAIN.NUM_GPUS
graph_funcs = MultiTowerOfflinePredictor(pred_config, list(
range(num_gpu))).get_predictors()
predictors = []
dataflows = []
for k in range(num_gpu):
predictors.append(
lambda img, pred=graph_funcs[k]: detect_one_image(img, pred))
dataflows.append(get_eval_dataflow(shard=k, num_shards=num_gpu))
if num_gpu > 1:
all_results = multithread_eval_coco(dataflows, predictors)
else:
all_results = eval_coco(dataflows[0], predictors[0])
with open(output_file, 'w') as f:
json.dump(all_results, f)
print_coco_metrics(output_file)
def predict(pred_func, input_file):
image_paths = []
if os.path.isdir(input_file):
for inp_file in os.listdir(input_file):
image_paths += [input_file + inp_file]
for image_path in image_paths:
(path1,file1)=os.path.split(image_path)
(filename1, filename2) = os.path.splitext(file1)
objectfile = open('results/%s.txt' % filename1,'w')
img = cv2.imread(image_path, cv2.IMREAD_COLOR)
start=time.time()
results = detect_one_image(img, pred_func)
end=time.time()
print(end-start)
final = draw_final_outputs(img, results,objectfile)
#viz = np.concatenate((img, final), axis=1)
viz=final
cv2.imwrite('outimages/%s.png'%filename1,viz)
def forward_protobuf(pred_func, output_folder, forward_dataset,
generic_images_folder, generic_images_pattern):
tf.gfile.MakeDirs(output_folder)
hypo_set = HypothesisSet()
with open(
"/work/merkelbach/datasets/dashcam_mining/tracking_results_raw/vid_01/eval/12-Mar-2018--19-25-00vayV/hypos_protobuf/hypotheses_500frames_1860hyp___work__pv182253__data__dashcam__dashcam_videos_frames__vid_01__frames_cropped__clip_005.hypset",
"rb") as f:
hypo_set.ParseFromString(f.read())
for hyp in hypo_set.hypotheses:
boxes = hyp.bounding_boxes_2D_with_timestamps
for t, box in boxes.items():
img_filename = "/work/merkelbach/datasets/dashcam_mining/videos/vid_01/frames_cropped/all_frames/video_0001_frames" + "%09d" % t + ".png"
print(img_filename, box)
img_val = cv2.imread(img_filename, cv2.IMREAD_COLOR)
assert config.PROVIDE_BOXES_AS_INPUT
input_boxes = np.array(
[[box.x0, box.y0, box.x0 + box.w, box.y0 + box.h]],
dtype=np.float32)
results = detect_one_image(img_val, pred_func, input_boxes)
print(results)
def generate_nii(pred_func, input_files):
arg = []
for idx, input_file in tqdm.tqdm(enumerate(input_files)):
boxes = collections.defaultdict(list)
masks = collections.defaultdict(list)
img = cv2.imread(input_file, cv2.IMREAD_COLOR)
results = detect_one_image(img, pred_func)
for res in results:
cl = config.CLASS_NAMES[res.class_id]
boxes[cl].append(list(res.box) + [res.score])
masks[cl].append(res.mask)
arg.append({
'image': input_file,
'boxes': boxes,
'masks': masks,
'im_info': (200, 200)
})
evaluate_mask_mean_iou(arg)
def offline_evaluate(pred_config, output_file):
num_gpu = cfg.TRAIN.NUM_GPUS
graph_funcs = MultiTowerOfflinePredictor(pred_config, list(
range(num_gpu))).get_predictors()
predictors = []
for k in range(num_gpu):
predictors.append(
lambda img, pred=graph_funcs[k]: detect_one_image(img, pred))
for dataset in cfg.DATA.VAL:
logger.info("Evaluating {} ...".format(dataset))
dataflows = [
get_eval_dataflow(dataset, shard=k, num_shards=num_gpu)
for k in range(num_gpu)
]
if num_gpu > 1:
all_results = multithread_eval_coco(dataflows, predictors)
else:
all_results = eval_coco(dataflows[0], predictors[0])
output = output_file + '-' + dataset
with open(output, 'w') as f:
json.dump(all_results, f)
print_coco_metrics(dataset, output)
def predict(pred_func, input_file):
img = cv2.imread(input_file, cv2.IMREAD_COLOR)
results = detect_one_image(img, pred_func)
final = draw_final_outputs(img, results)
viz = np.concatenate((img, final), axis=1)
tpviz.interactive_imshow(viz)
def forward(pred_func,
output_folder,
forward_dataset,
generic_images_folder=None,
generic_images_pattern=None):
if forward_dataset.lower() == "davis":
# name = "/home/luiten/vision/PReMVOS/home_data/%s/images/*"%config.DAVIS_NAME
# imgs = glob.glob(name)
# seq_idx = -2
# name = config.DAVIS_NAME
# imgs = glob.glob(name+"/*/*")
# seq_idx = -2
name = config.DAVIS_NAME
pre = '/'.join(name.split('/')[:-1])
print(name)
imgs = []
f = open(name, 'r')
while True:
x = f.readline()
x = x.rstrip()
if not x: break
print(pre + '/' + x)
imgs = imgs + glob.glob(pre + '/' + x + '/*')
# for file in open(name):
# imgs = imgs + glob.glob(file+'/*')
print(imgs)
seq_idx = -2
elif forward_dataset.lower() == "oxford":
imgs = [
"/fastwork/" + os.environ["USER"] + "/mywork/" + x.strip() for x in
open("/home/" + os.environ["USER"] +
"/vision/TrackingAnnotationTool/to_annotate_single_imgs.txt").
readlines() if len(x.strip()) > 0
]
seq_idx = -4
elif forward_dataset.lower() == "kitti_tracking":
imgs = glob.glob(
"/fastwork/" + os.environ["USER"] +
"/mywork/data/kitti_training_minimum/training/image_02/*/*.png")
seq_idx = -2
elif forward_dataset.lower() == "generic":
generic_images_pattern = generic_images_pattern.replace("//", "/")
assert generic_images_folder is not None, "For a generic dataset, a data folder must be given."
assert generic_images_pattern is not None, "For a generic dataset, an image pattern must be given."
# Find the images in the folder
imgs = sorted(
glob.glob(
os.path.join(generic_images_folder, generic_images_pattern)))
# The sequence (if there is any) is assumed to be the folder the images are in
#seq_idx = -1 - generic_images_pattern.count("/")
seq_idx = None
elif forward_dataset.lower() == "protobuf":
forward_protobuf(pred_func, output_folder, forward_dataset,
generic_images_folder, generic_images_pattern)
else:
assert False, ("Unknown dataset", forward_dataset)
tf.gfile.MakeDirs(output_folder)
n_total = len(imgs)
for idx, img in enumerate(imgs):
print(idx, "/", n_total)
if forward_dataset.lower() == "generic":
seq = "/".join(
img.replace(generic_images_folder, "").split("/")[:-1])
seq_folder = output_folder + "/" + seq
else:
seq = img.split("/")[seq_idx]
seq_folder = output_folder + "/" + seq
img_filename = img.split("/")[-1]
tf.gfile.MakeDirs(seq_folder)
output_filename = seq_folder + "/" + img_filename.replace(
".jpg", ".json").replace(".png", ".json")
if os.path.isfile(output_filename):
print("skipping", output_filename, "because it already exists")
continue
else:
print(output_filename)
img_val = cv2.imread(img, cv2.IMREAD_COLOR)
assert img_val is not None, ("unable to load", img)
if config.PROVIDE_BOXES_AS_INPUT:
# TODO get some reasonable boxes (not sure which format)
input_boxes = np.array([[202.468384, 566.15271, 999, 800],
[0, 0, 100, 100], [100, 100, 200, 200]],
np.float32)
results = detect_one_image(img_val, pred_func, input_boxes)
else:
results = detect_one_image(img_val, pred_func)
print(len(results))
# store visualization (slow + optional)
if args.forward_visualization:
final = draw_final_outputs(img_val, results)
viz = np.concatenate((img_val, final), axis=1)
#tpviz.interactive_imshow(viz)
viz_output_filename = seq_folder + "/" + img_filename.replace(
".png", ".jpg")
cv2.imwrite(viz_output_filename, viz)
# store as json
results_json = convert_results_to_json(results, idx)
with open(output_filename, 'w') as f:
json.dump(results_json, f)
def offline_evaluate(pred_func, output_file):
df = get_test_dataflow()
all_results = eval_on_dataflow(
df, lambda img: detect_one_image(img, pred_func))
print(all_results)
def predict(pred_func, input_file):
for frame_name in os.listdir(input_file):
intensity_file = pd.read_csv(os.path.join(intensity_path, frame_name))
pts_file = pd.read_csv(os.path.join(pts_path, frame_name))
file_name = str.split(frame_name, '.')[0]
pts = pd.read_csv(os.path.join(pts_path, file_name + '.csv'),
header=None)
pts = np.array(pts)
out = np.zeros((len(pts), 1), dtype=np.int8)
start = time.time()
PointCloud = np.hstack((pts_file, intensity_file))
img = makeBVFeature(PointCloud, size_ROI, size_cell)
results = detect_one_image(img, pred_func)
#objectfile=open('results1/%s.txt'%file_name,'w')
pts[:, 0] = maxX - pts[:, 0] # 坐标系统一为左上角
pts[:, 1] = maxY - pts[:, 1]
# PointCloud_grid = np.copy(PointCloud )
pts[:, 0] = np.int_(np.floor(pts[:, 0] /
size_cell)) # np.floor 返回不大于输入参数的最大整数；
pts[:, 1] = np.int_(np.floor(pts[:, 1] / size_cell))
for r in results:
label = cfg.DATA.CLASS_NAMES[r.class_id][0:3]
#if r.score<0.4:
#label=0
indx = np.where((pts[:, 1] > int(r[0][0]))
& (pts[:, 1] < int(r[0][2]))
& (pts[:, 0] < int(r[0][3]))
& (pts[:, 0] > int(r[0][1])))
if label == 'cyc':
if r.score < 0.6:
label = 0
else:
label = 1
# indx = np.where(
# (pts[:, 1] > int(r[0][0])) & (pts[:, 1] < int(r[0][2])) & (pts[:, 0] < int(r[0][3])) & (pts[:, 0] > int(r[0][1]))& (pts[:, 2] <-0.3))
elif label == 'tri':
label = 2
# indx = np.where(
# (pts[:, 1] > int(r[0][0])) & (pts[:, 1] < int(r[0][2])) & (pts[:, 0] < int(r[0][3])) & (
# pts[:, 0] > int(r[0][1])) & (pts[:, 2] < -0.3))
elif label == 'sma':
label = 3
# indx = np.where(
# (pts[:, 1] > int(r[0][0])) & (pts[:, 1] < int(r[0][2])) & (pts[:, 0] < int(r[0][3])) & (
# pts[:, 0] > int(r[0][1])) & (pts[:, 2] < 0))
elif label == 'big':
label = 4
# indx = np.where(
# (pts[:, 1] > int(r[0][0])) & (pts[:, 1] < int(r[0][2])) & (pts[:, 0] < int(r[0][3])) & (
# pts[:, 0] > int(r[0][1])) )
elif label == 'ped':
if r.score < 0.4:
label = 0
else:
label = 5
#label = 5
# indx = np.where(
# (pts[:, 1] > int(r[0][0])) & (pts[:, 1] < int(r[0][2])) & (pts[:, 0] < int(r[0][3])) & (
# pts[:, 0] > int(r[0][1])) & (pts[:, 2] < -0.1))
elif label == 'cro':
label = 6
# indx = np.where(
# (pts[:, 1] > int(r[0][0])) & (pts[:, 1] < int(r[0][2])) & (pts[:, 0] < int(r[0][3])) & (
# pts[:, 0] > int(r[0][1])) & (pts[:, 2] < -0.1))
elif label == 'unk':
if r.score < 0.3:
label = 0
# else:
# label=7
# indx = np.where(
# (pts[:, 1] > int(r[0][0])) & (pts[:, 1] < int(r[0][2])) & (pts[:, 0] < int(r[0][3])) & (
# pts[:, 0] > int(r[0][1])) )
label = 7
out[indx] = label
end = time.time()
np.savetxt(os.path.join(upload_path, file_name + '.csv'),
out,
delimiter=',',
fmt='%d')
print(end - start)
# autotune is too slow for inference
os.environ['TF_CUDNN_USE_AUTOTUNE'] = '0'
assert args.load
print_config()
if args.evaluate:
pred = OfflinePredictor(
PredictConfig(model=Model(),
session_init=get_model_loader(args.load),
input_names=['image'],
output_names=get_model_output_names()))
df = get_test_dataflow(add_mask=True)
df.reset_state()
all_results, local_score = eval_on_dataflow(
df, lambda img: detect_one_image(img, pred))
print("F2 Score: ", local_score)
elif args.predict:
imgs = Detection.load_many(
config.BASEDIR, config.TEST_DATASET,
add_gt=False) # to load the class names into caches
# filter with zero-ship
imgs = [(img['image_data'], img['id']) for img in imgs]
pred = OfflinePredictor(
PredictConfig(model=Model(),
session_init=get_model_loader(args.load),
input_names=['image'],
output_names=get_model_output_names()))
predict_many(pred, imgs)
else:
def _build_coco_predictor(self, idx):
graph_func = self.trainer.get_predictor(self._in_names,
self._out_names,
device=idx)
return lambda img: detect_one_image(img, graph_func)
def predict_many(pred_func, input_files):
import time
import pandas as pd
with open("./id_no_shop_384_2.pkl", "rb") as f:
id_no_ship = pickle.load(f)
#fid = open('result.csv','w+')
#fid.write('ImageId,EncodedPixels\n')
ship_list_dict = []
with tqdm.tqdm(total=len(input_files)) as pbar:
for idx, imgData in enumerate(input_files):
img = cv2.imread(imgData[0])
filename = imgData[1]
ImageId = filename
if ImageId in id_no_ship:
ship_list_dict.append({
'ImageId': ImageId,
'EncodedPixels': np.nan
})
pbar.update()
continue
#s = time.time()
results = detect_one_image(img.copy(), pred_func)
mask_instances = [r.mask for r in results]
score_instances = [r.score for r in results]
#mask_whole = detect_one_image(img.copy(), pred_func)
"""
#print(time.time() - s)
if (len(results) == 0):
# no detection in image
result_one_line = ImageId+','+ "" +'\n'
ship_list_dict.append({'ImageId':ImageId,'EncodedPixels':np.nan})
#fid.write(result_one_line)
pbar.update()
continue
r = {}
r['class_ids'] = []
r['scores'] = []
r['masks'] = []
for det in results:
if np.count_nonzero(det.mask) > 0:
r['class_ids'].append(det.class_id)
r['scores'].append(det.score)
r['masks'].append(det.mask)
if len(r['masks']) == 0:
ship_list_dict.append({'ImageId':ImageId,'EncodedPixels':np.nan})
print('no_mask')
pbar.update()
continue
r['masks'] = np.array(r['masks']) # n, h, w
#r['masks'] = np.transpose(r['masks'], [1,2,0])
ImageId = filename
#print(filename, r['masks'].shape)
LabelId = r['class_ids']
mask_whole = np.zeros((img.shape[0], img.shape[1]))
#mask_clean = clean_overlap_instance(r['masks'], r['scores'])
mask_clean = r['masks']
for i in range(mask_clean.shape[0]):
mask_whole[mask_clean[i] > 0] = 1
#EncodedPixels = rle_encoding(mask_clean[i])
#ship_list_dict.append({'ImageId':ImageId,'EncodedPixels':EncodedPixels})
"""
masks = clean_overlap_instance(mask_instances, score_instances)
#masks = split_mask(mask_whole)
if len(masks) == 0:
ship_list_dict.append({
'ImageId': ImageId,
'EncodedPixels': np.nan
})
print('no_mask!!!')
pbar.update()
continue
for mask in masks:
ship_list_dict.append({
'ImageId': ImageId,
'EncodedPixels': rle_encoding(mask)
})
#if idx < 30:
# cv2.imwrite(os.path.join("output", filename), mask_whole*255)
pbar.update()
#fid.close()
pred_df = pd.DataFrame(ship_list_dict)
pred_df = pred_df[['ImageId', 'EncodedPixels']]
pred_df.to_csv('submission.csv', index=False)
print('done!')
def forward(pred_func,
output_folder,
forward_dataset,
generic_images_folder=None,
generic_images_pattern=None):
if forward_dataset.lower() == "davis":
imgs = glob.glob("/fastwork/" + os.environ["USER"] +
"/mywork/data/DAVIS2017/JPEGImages/480p/*/00000.jpg")
seq_idx = -2
elif forward_dataset.lower() == "oxford":
mywork_dir = "/fastwork/" + os.environ["USER"] + "/mywork/"
if args.mywork_folder is not None:
mywork_dir = args.mywork_folder
imgs = [
os.path.join(mywork_dir, x.strip()) for x in
open("/home/" + os.environ["USER"] +
"/vision/TrackingAnnotationTool/to_annotate_single_imgs.txt").
readlines() if len(x.strip()) > 0
]
seq_idx = -4
elif forward_dataset.lower() == "kitti_tracking":
imgs = glob.glob(
"/fastwork/" + os.environ["USER"] +
"/mywork/data/kitti_training_minimum/training/image_02/*/*.png")
seq_idx = -2
elif forward_dataset.lower() == "generic":
generic_images_pattern = generic_images_pattern.replace("//", "/")
assert generic_images_folder is not None, "For a generic dataset, a data folder must be given."
assert generic_images_pattern is not None, "For a generic dataset, an image pattern must be given."
# Find the images in the folder
imgs = sorted(
glob.glob(
os.path.join(generic_images_folder, generic_images_pattern)))
print("Num. images: %d" % len(imgs))
seq_idx = None
else:
assert False, ("Unknown dataset", forward_dataset)
tf.gfile.MakeDirs(output_folder)
n_total = len(imgs)
t0 = time.time()
for idx, img in enumerate(imgs):
print(idx, "/", n_total)
if forward_dataset.lower() == "generic":
seq = "/".join(
img.replace(generic_images_folder, "").split("/")[:-1])
seq_folder = output_folder + "/" + seq
else:
seq = img.split("/")[seq_idx]
seq_folder = output_folder + "/" + seq
img_filename = img.split("/")[-1]
tf.gfile.MakeDirs(seq_folder)
output_filename = seq_folder + "/" + img_filename.replace(
".jpg", ".json").replace(".png", ".json")
output_filename_embeddings = seq_folder + "/" + img_filename.replace(
".jpg", ".npz").replace(".png", ".npz")
if os.path.isfile(output_filename):
print("skipping", output_filename, "because it already exists")
continue
else:
print(output_filename)
img_val = cv2.imread(img, cv2.IMREAD_COLOR)
assert img_val is not None, ("unable to load", img)
if config.PROVIDE_BOXES_AS_INPUT:
# TODO get some reasonable boxes (not sure which format)
input_boxes = np.array([[202.468384, 566.15271, 999, 800],
[0, 0, 100, 100], [100, 100, 200, 200]],
np.float32)
results = detect_one_image(img_val, pred_func, input_boxes)
else:
t0_det = time.time()
results = detect_one_image(img_val, pred_func)
t1_det = time.time()
print(len(results))
# store visualization (slow + optional)
if args.forward_visualization:
final = draw_final_outputs(img_val, results)
viz = np.concatenate((img_val, final), axis=1)
#tpviz.interactive_imshow(viz)
viz_output_filename = seq_folder + "/" + img_filename.replace(
".png", ".jpg")
cv2.imwrite(viz_output_filename, viz)
# store as json
t0_json = time.time()
results_json = convert_results_to_json(results, idx)
t1_json = time.time()
print('Json conv. time: %f s' % (t1_json - t0_json))
with open(output_filename, 'w') as f:
#json.dump(results_json, f, sort_keys=True, indent=4, separators=(',', ': '))
json.dump(results_json, f) # Non-pretty
M_emb = convert_embeddings_to_binary(results)
t0_cmp_npz = time.time()
np.savez_compressed(output_filename_embeddings,
a=M_emb.astype(np.float16))
t1_cmp_npz = time.time()
t1 = time.time()
time_elapsed = t1 - t0
def _build_coco_predictor(self, idx):
graph_func = self.trainer.get_predictor(self._in_names, self._out_names, device=idx)
return lambda img: detect_one_image(img, graph_func)