Can I obtain a simil file like the file from the command --save-txt of Yolov5?

[claudio9russo7]

Hello
i didn't find anything but maybe i'm looking wrong, is it possible to get a txt file like the one you get by doing the inference with yolov5 on the pictures?

[fcakyon]

@claudio9russo7 you can export coco result json file by providing dataset json to the predict command: https://github.com/obss/sahi/issues/356

[claudio9russo7]

@fcakyon Ok, thank you. This is true only if they are already images with annotation or not?

[claudio9russo7]

I found similar the --export_pickle command.

[fcakyon]

Yes, that arg is relevant 👍 @claudio9russo7 does --export_pickle solve your issue?

[claudio9russo7]

Yep, solve my issue, I just want to suggest adding a possible choice of export format like .json or .txt :D

[fcakyon]

@claudio9russo7 noted it, will look into that feature when I find some time 🚀

[ErToBar]

here is my version:;
import os
import pickle
import re
from PIL import Image

Function to extract details from object prediction

def extract_details_from_prediction(obj_pred):
bbox_values = re.search(r'<((.*?))', str(obj_pred.bbox)).group(1)
bbox = list(map(float, bbox_values.split(',')))

score = float(re.search(r'value: (.*?)[,>]', str(obj_pred.score)).group(1))

category_values = re.search(r'id: (.*?),', str(obj_pred.category)).group(1)
category = int(category_values.strip())

return bbox, score, category

Path of the visuals directory

visuals_path = ""

Path of the pickles directory

pickles_path = ""

Output path

output_path = ""

Iterate over all images in the visuals directory

for image_file_name in os.listdir(visuals_path):
image_path = os.path.join(visuals_path, image_file_name)

# Corresponding pickle file name and path
pickle_file_name = os.path.splitext(image_file_name)[0] + ".pickle"
pickle_file_path = os.path.join(pickles_path, pickle_file_name)

# Check if the pickle file exists
if not os.path.exists(pickle_file_path):
    print(f"No corresponding pickle file found for image: {image_file_name}")
    continue

# Load the pickle file data
with open(pickle_file_path, 'rb') as f:
    data = pickle.load(f)

# Get the dimensions of the image
image = Image.open(image_path)
image_width, image_height = image.size

# Iterate over predictions and write to file
for obj_pred in data:
    # Extract details from prediction
    bbox, score, category = extract_details_from_prediction(obj_pred)

    # Calculate the center of the box, width and height
    bbox_center_x = (bbox[0] + bbox[2]) / 2.0
    bbox_center_y = (bbox[1] + bbox[3]) / 2.0
    bbox_width = bbox[2] - bbox[0]
    bbox_height = bbox[3] - bbox[1]

    # Normalize the values
    bbox_center_x /= image_width
    bbox_center_y /= image_height
    bbox_width /= image_width
    bbox_height /= image_height

    # Save to the output file
    output_file_name = os.path.splitext(image_file_name)[0] + ".txt"
    output_file_path = os.path.join(output_path, output_file_name)

    with open(output_file_path, "a") as out_f:  # use "a" to append to file for multiple predictions
        out_f.write(f"{category} {bbox_center_x} {bbox_center_y} {bbox_width} {bbox_height} {score}\n")

[brianbattaile]

As a beginner, I tried to implement the above code but it threw all kinds of errors. What exactly should this code look like as a whole unit and at what points does a user need to fill in paths etc.? I get I need to add a visuals_path, pickles_path and output_path. What else am I missing? I tried this without success

import os
import pickle
import re
from PIL import Image

def extract_details_from_prediction(obj_pred):
bbox_values = re.search(r'<((.*?))', str(obj_pred.bbox)).group(1)
bbox = list(map(float, bbox_values.split(',')))

core = float(re.search(r'value: (.*?)[,>]', str(obj_pred.score)).group(1))

category_values = re.search(r'id: (.*?),', str(obj_pred.category)).group(1)
category = int(category_values.strip())

return bbox, score, category

visuals_path = r"C:\Users\Green Sturgeon\AI_Project\Test\runs\predict\exp\visuals"
pickles_path = r"C:\Users\Green Sturgeon\AI_Project\Test\runs\predict\exp\pickles"
output_path = r"C:\Users\Green Sturgeon\AI_Project\Test\runs\predict\exp"

for image_file_name in os.listdir(visuals_path):
image_path = os.path.join(visuals_path, image_file_name)

# Corresponding pickle file name and path
pickle_file_name = os.path.splitext(image_file_name)[0] + ".pickle"
pickle_file_path = os.path.join(pickles_path, pickle_file_name)

# Check if the pickle file exists
if not os.path.exists(pickle_file_path):
    print(f"No corresponding pickle file found for image: {image_file_name}")
    continue

# Load the pickle file data
with open(pickle_file_path, 'rb') as f:
    data = pickle.load(f)

# Get the dimensions of the image
image = Image.open(image_path)
image_width, image_height = image.size

# Iterate over predictions and write to file
for obj_pred in data:
    # Extract details from prediction
    bbox, score, category = extract_details_from_prediction(obj_pred)

    # Calculate the center of the box, width and height
    bbox_center_x = (bbox[0] + bbox[2]) / 2.0
    bbox_center_y = (bbox[1] + bbox[3]) / 2.0
    bbox_width = bbox[2] - bbox[0]
    bbox_height = bbox[3] - bbox[1]

    # Normalize the values
    bbox_center_x /= image_width
    bbox_center_y /= image_height
    bbox_width /= image_width
    bbox_height /= image_height

    # Save to the output file
    output_file_name = os.path.splitext(image_file_name)[0] + ".txt"
    output_file_path = os.path.join(output_path, output_file_name)

    with open(output_file_path, "a") as out_f:  # use "a" to append to file for multiple predictions
        out_f.write(f"{category} {bbox_center_x} {bbox_center_y} {bbox_width} {bbox_height} {score}\n")

[brianbattaile]

Dang, I have no idea why the code section formatted the way it did in the previous post...let me try again.

import os
import pickle
import re
from PIL import Image

def extract_details_from_prediction(obj_pred):
bbox_values = re.search(r'<((.*?))', str(obj_pred.bbox)).group(1)
bbox = list(map(float, bbox_values.split(',')))

core = float(re.search(r'value: (.*?)[,>]', str(obj_pred.score)).group(1))

category_values = re.search(r'id: (.*?),', str(obj_pred.category)).group(1)
category = int(category_values.strip())

return bbox, score, category

visuals_path = r"C:\Users\Green Sturgeon\AI_Project\Test\runs\predict\exp\visuals"
pickles_path = r"C:\Users\Green Sturgeon\AI_Project\Test\runs\predict\exp\pickles"
output_path = r"C:\Users\Green Sturgeon\AI_Project\Test\runs\predict\exp"

for image_file_name in os.listdir(visuals_path):
image_path = os.path.join(visuals_path, image_file_name)

# Corresponding pickle file name and path
pickle_file_name = os.path.splitext(image_file_name)[0] + ".pickle"
pickle_file_path = os.path.join(pickles_path, pickle_file_name)

# Check if the pickle file exists
if not os.path.exists(pickle_file_path):
    print(f"No corresponding pickle file found for image: {image_file_name}")
    continue

# Load the pickle file data
with open(pickle_file_path, 'rb') as f:
    data = pickle.load(f)

# Get the dimensions of the image
image = Image.open(image_path)
image_width, image_height = image.size

# Iterate over predictions and write to file
for obj_pred in data:
    # Extract details from prediction
    bbox, score, category = extract_details_from_prediction(obj_pred)

    # Calculate the center of the box, width and height
    bbox_center_x = (bbox[0] + bbox[2]) / 2.0
    bbox_center_y = (bbox[1] + bbox[3]) / 2.0
    bbox_width = bbox[2] - bbox[0]
    bbox_height = bbox[3] - bbox[1]

    # Normalize the values
    bbox_center_x /= image_width
    bbox_center_y /= image_height
    bbox_width /= image_width
    bbox_height /= image_height

    # Save to the output file
    output_file_name = os.path.splitext(image_file_name)[0] + ".txt"
    output_file_path = os.path.join(output_path, output_file_name)

    with open(output_file_path, "a") as out_f:  # use "a" to append to file for multiple predictions
        out_f.write(f"{category} {bbox_center_x} {bbox_center_y} {bbox_width} {bbox_height} {score}\n")

[brianbattaile]

Still wrong, I'm clueless.

[andressrodrl]

https://github.com/andresinsitu/sahi_custom/tree/export_txt

Ive modified sahi so you can save the predictions txt in yolo format. You can use it like this:

Add the argument export_yolo=True or the argument export_yolo_conf=True to the predict function. The first one only saves xywh and the second one also saves confidence so xywhc . Results will be saved to a "labels" folder

[brianbattaile]

How can I install your version into my local python environment so I can use it instead of the standard SAHI version I already have installed there?

…

On Wed, Nov 15, 2023 at 1:37 AM andresinsitu ***@***.***> wrote: https://github.com/andresinsitu/sahi_custom Ive modified sahi so you can save the predictions txt in yolo format. You can use it like this: Add the argument export_yolo=True or the argument export_yolo_conf=True to the predict function. The first one only saves xywh and the second one also saves confidence so xywhc . Results will be saved to a "labels" folder — Reply to this email directly, view it on GitHub <#662 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AZIHUOS62UOEW6FPB35L6V3YESENJAVCNFSM6AAAAAAZYIS5G6VHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM3TKNZUHA4TO> . You are receiving this because you commented.Message ID: ***@***.***>