Hello, I hope you are doing well. The yolov5 and yolov8 save the output text file with the Class label and bounding box value. I want to save time when the object is detected and the class name is not labeled. Can any Body Help me with that
@Hamad_Younis, you can use the time and/or datetime modules from the standard python library.
Import the library and log a record of the current time when the detection was made.
Here’s my suggestions after looking at the YOLOv5 repo’s code
For YOLOv5, use the arguments in detect.py to create the txt file: yolov5/detect.py at master · ultralytics/yolov5 · GitHub
Utilize the --save-txt flag to create a txt file of your detections, and include the --save-conf flag to include the confidence level for the detctions
To include the time, modify the detect.py file to include a function for extracting the current time, and creating a record for it in string format:
- Short example:
import time
# Initialize timer
t1 = time.time()
# Run inference
results = model(img)
# Calculate elapsed time
t2 = time.time()
dt = t2 - t1
# Print detection time
print(f"Detection time: {dt:.4f} seconds")
Press enter at the end of line 81 to create a new line, and then enter this in the new line:
save_time=False
- Example:
def run(
weights=ROOT / 'yolov5s.pt', # model path or triton URL
source=ROOT / 'data/images', # file/dir/URL/glob/screen/0(webcam)
data=ROOT / 'data/coco128.yaml', # dataset.yaml path
imgsz=(640, 640), # inference size (height, width)
conf_thres=0.25, # confidence threshold
iou_thres=0.45, # NMS IOU threshold
max_det=1000, # maximum detections per image
device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu
view_img=False, # show results
save_txt=False, # save results to *.txt
save_conf=False, # save confidences in --save-txt labels
save_crop=False, # save cropped prediction boxes
nosave=False, # do not save images/videos
classes=None, # filter by class: --class 0, or --class 0 2 3
agnostic_nms=False, # class-agnostic NMS
augment=False, # augmented inference
visualize=False, # visualize features
update=False, # update all models
project=ROOT / 'runs/detect', # save results to project/name
name='exp', # save results to project/name
exist_ok=False, # existing project/name ok, do not increment
line_thickness=3, # bounding box thickness (pixels)
hide_labels=False, # hide labels
hide_conf=False, # hide confidences
half=False, # use FP16 half-precision inference
dnn=False, # use OpenCV DNN for ONNX inference
vid_stride=1, # video frame-rate stride
save_time=False # choose to save timestamp for the detection
):
Then, add a new line between line 247 and line 248 that reads:
parser.add_argument('--save_time', action='store_true', help='Save the timestamp of each detection')
- This will create the argument to pass when running
detect.py, to store the timestamps
Line 63 is where you’ll want to start making the updates for the timestamp logic:
Add these new lines after line 128 to create the logic for saving the detection timestamp records:
for detected_object in pred:
if save_time:
timestamp = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime())
detection_data.append({'object': detected_object, 'timestamp': timestamp})
else:
detection_data.append({'object': detected_object})
Finally, modify the code, beginning with line 168 to include the detection records, if the --save_time flag is True:
time_index = 0
if save_time: # save timestamp to file
f.write(("%g " * len(line)).rstrip() % line + f", {detection_data[time_index]}" + "\n")
time_index+=1
else:
f.write(('%g ' * len(line)).rstrip() % line + '\n')
For YOLOv8, you’ll want to make the modifications in the predict.py file corresponding to your model type:
In the case of YOLOv8, try adding your new argument to the model.py file, and within the YOLO class:
After you’ve given implementation a try, please reply back to confirm if this works, or if you have any errors while running your code.
@Mohamed Thanks for your great reply, But I am facing an issue with this.
Can you share the final in which you have edit the code detect.py file for yolov5 and predict.py for yolov8
@Hamad_Younis can you provide more specifics on your error?
Such as the exact error, and a screenshot of it?
I provided the code/suggestions based on my reading of the repo, so I don’t have the direct files handy.
I’ve also only looked at the YOLOv5 code and provided suggested code for that. I haven’t yet had a chance to write full code for how to do this for YOLOv8.
I’ll try to do so for YOLOv8 when I have time, but if you’re pressed for time, it may be better to use the formatting from the YOLOv5 version of it, once its working, and update the YOLOv8 code.
@Mohamed I am unable to do that for yolov5, Can you please only modified the code for YOLOv5, not for YOLOv8.
If you can share detect.py for yolov5 with edit change i will be very thankful to you
@Mohamed Can you please help me i am kind of stuck with it for the past few days.
if you can modify detect.py code for YOLOv5 I will be very thankful to you
@Mohamed
Can you please help me i am kind of stuck with it for the past few days.
if you can modify detect.py code for YOLOv5 I will be very thankful to you
@Mohamed Error
Traceback (most recent call last):
File “/content/yolov5/detect.py”, line 277, in
main(opt)
File “/content/yolov5/detect.py”, line 272, in main
run(**vars(opt))
File “/usr/local/lib/python3.9/dist-packages/torch/autograd/grad_mode.py”, line 27, in decorate_context
return func(*args, **kwargs)
File “/content/yolov5/detect.py”, line 183, in run
f.write(('%g ’ * len(line)).rstrip() % line + ‘\n’)
UnboundLocalError: local variable ‘f’ referenced before assignment
terminate called without an active exception
@Mohamed It saving the timestamp like below
33 0.808984 0.968843 0.0539063 0.0623145, {‘timestamp’: ‘2023-03-25 07:02:08’}
But I want the time of the Video no the current time
1 Like
Adding this here so everything remains on the same thread (this one), since it’s the same inquiry, rather than a new one:
@Hamad_Younis Tested and looks like I got it to work with my face detection model on Roboflow:
To run it with YOLOv5 or YOLOv8, upload your model weights to Roboflow, and use the code below:
- I’m giving the Roboflow example because it will allow you to use the same code, with no extra configuration other than ensuring the workspace ID, project ID, model ID/version number, and Private API Key point to the model that holds the weights for the architecture you desire to run inference with (YOLOv5 or YOLOv8)
- Python Package - Roboflow | GitHub - roboflow/roboflow-python: Interface with Roboflow in Python
- The fps for processing will be determined by your hardware. I ran mine with CPU only, and no GPU.
import time
import datetime
import cv2
from roboflow import Roboflow
rf = Roboflow(api_key="PRIVATE_API_KEY")
project = rf.workspace("WORKSPACE_ID").project("PROJECT_ID")
model = project.version(VERSION_NUMBER).model # the version number is an integer value
# create our dictionary to save the class names and index/numeric value for the class
model_classes = project.classes
label_dict = {}
for i, label_name in enumerate(model_classes):
label_dict[label_name] = f"{i}"
cap = cv2.VideoCapture("REPLACE_WITH_PATH_TO_VIDEO")
# Get the total number of frames in the video
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
# Get the frame rate of the video
fps = int(cap.get(cv2.CAP_PROP_FPS))
# Calculate the duration of the video
duration = total_frames / fps
# Convert the video length to minutes and seconds
vid_minutes = duration // 60
vid_seconds = duration % 60
vid_length = datetime.time(minute=int(vid_minutes), second=int(vid_seconds))
print(f"The video is {duration} seconds long", '\n')
# process the video
while True:
ret, frame = cap.read()
t0 = time.time()
if not ret:
break
# Convert the frame from BGR to RGB
rgb_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
# Save the RGB frame as a JPEG file
img = cv2.imwrite('temp.jpg', rgb_frame)
# Detect objects in the frame
detected_objects = model.predict('temp.jpg', confidence=40, overlap=30).json()
# main bounding box coordinates from JSON response object
# https://docs.roboflow.com/inference/hosted-api#response-object-format
for bounding_box in detected_objects['predictions']:
print(detected_objects)
print(bounding_box)
xmin = bounding_box['x'] - bounding_box['width'] / 2
xmax = bounding_box['x'] + bounding_box['width'] / 2
ymin = bounding_box['y'] - bounding_box['height'] / 2
ymax = bounding_box['y'] + bounding_box['height'] / 2
confidence = bounding_box['confidence']
name = bounding_box['class']
# Find the label name corresponding corresponding to the class ID index value
for key, value in label_dict.items():
if key == name:
object_class = value
# position coordinates: start = (x0, y0), end = (x1, y1)
# color = RGB-value for bounding box color, (0,0,0) is "black"
# thickness = stroke width/thickness of bounding box
# draw and place bounding boxes
start_point = (int(xmin), int(ymin))
end_point = (int(xmax), int(ymax))
cv2.rectangle(frame, start_point, end_point, color=(0,0,0), thickness=2)
# calculate seconds per frame we are running inference
t = time.time()-t0
# Calculate the remaining duration of the video
remaining_duration = total_frames / fps
# Convert the remaining video length to minutes and seconds
remaining_minutes = remaining_duration // 60
remaining_seconds = remaining_duration % 60
# print the remaining video duration
remaining_duration = datetime.time(minute=int(remaining_minutes), second=int(remaining_seconds))
elapsed_time = datetime.datetime.combine(
datetime.date.min, vid_length) - datetime.datetime.combine(datetime.date.min, remaining_duration)
time_remaining = datetime.datetime.combine(
datetime.date.min, remaining_duration) - datetime.datetime.combine(datetime.date.min, vid_length)
if remaining_duration < vid_length:
reported_elapsed_dt = datetime.datetime(2000, 1, 1) + elapsed_time
timestamp = f"{reported_elapsed_dt.strftime('%M:%S')}"
# Draw the timestamp on the frame
cv2.putText(frame, f"{timestamp}", (10, 50), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
# If object(s) detected in the frame, create a variable to store their labels with the timestamp
if xmin:
label_with_timestamp = f"{xmin} {ymin} {xmax} {ymax} {confidence:.04f} {object_class} {name} {timestamp}"
else:
reported_elapsed_dt = datetime.datetime(2000, 1, 1) + elapsed_time
timestamp = f"{reported_elapsed_dt.strftime('%M:%S')}"
# Draw the timestamp on the frame
cv2.putText(frame, f"{timestamp}", (10, 50), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
# If object(s) detected in the frame, create a variable to store their labels with the timestamp
if xmin:
label_with_timestamp = f"{xmin} {ymin} {xmax} {ymax} {confidence:.04f} {object_class} {name} {timestamp}"
with open(f'inference_results.txt', 'a') as f:
f.write(label_with_timestamp + '\n')
# Logging purposes
print('\n', f'Processing {1/t} Frames Per Second (FPS): ', '\n')
print('\n', 'Written to txt file: ', label_with_timestamp, '\n')
cv2.imshow("frame", frame)
total_frames-=1
if cv2.waitKey(1) == ord("q"):
break
cap.release()
cv2.destroyAllWindows()
^ scroll down with your mouse hovered over the code to see/copy all of it
@Mohamed Thanks for your work. I will check it in a while. The Detected Video also save in folder as detect.py save the annotated video?
You can save the video with detections if you also use the cv2.VideoWriter() method from the OpenCV library:
You’ll just be saving the frames from the frame variable. Include the method for writing to the actual output stream just before the cv2.imshow() line in the code I provided above.