Since I was interested in the dataset I naturally explored the annotated and additional data that can be reached by the provided links in the README.md. However, as exploring the additional data provided by the link http://go.mytum.de/928009 I realized one can open more directories than probably anticipated. In short by going up the directory path one can reach a directory called Studien containing files like Kennzahlensysteme_in_der_Automobillogistik.pdf. I am not providing the download link here as I am unsure whether this was intended but I wanted to raise awareness as it doesn't seem related to the LOCO dataset.

Hi,
Thank you very much for your interesting work and for publicly releasing your dataset.
Do you have any plans to release the checkpoints of trained Yolo-V4 & Fater-RCNN detection methods?

Thanks

I was trying to use this dataset with pytorch and for ease of use I decided to use the given CocoDetection class (https://pytorch.org/vision/stable/_modules/torchvision/datasets/coco.html#CocoDetection.__getitem__), which was created for datasets in the coco format. Unfortunately, I quickly realized the dataset is not in the expected form for this class. The class does not accept seperate file paths for each image, it expects all images to be in one folder.

I checked if all images have a unique name. Indeed they have. So I decided to move all images to one folder.
Here is a small python script to do that:

import shutil

# Set the directory you want to start from
original_dir_path = '/notebooks/dataset/loco_old'
copy_dir_path = '/notebooks/dataset/loco_new'
destination_dir = '/notebooks/dataset/loco_new/images/'

# Supported image extensions
image_extensions = {".jpg", ".jpeg"}

# Step 1: Make a copy of the original folder
shutil.copytree(original_dir_path, copy_dir_path)

# Step 2: Create a destination directory inside the copied folder if not exists
os.makedirs(os.path.join(copy_dir_path, destination_dir), exist_ok=True)

# Step 3: Traverse through the directory including subdirectories of the copied folder
for dir_name, sub_dir_list, file_list in os.walk(copy_dir_path):
    for file_name in file_list:
        # Check if the file is an image based on its extension
        if any(file_name.lower().endswith(ext) for ext in image_extensions):
            # Construct full file path
            source = os.path.join(dir_name, file_name)
            destination = os.path.join(copy_dir_path, destination_dir, file_name)
            # Move the file to the destination directory
            shutil.move(source, destination)

print("All images have been moved to the destination directory.")

Hope it helps some people. It wasted quite some time for me to realize what the error is and to fix it.

I downloaded the dataset.zip. However, the the described json file is not contained in it.

Annotations are stored in COCO format under rgb/loco-all-v1.json. For ease of use, we also provide seperate annotation files for each subset.

Within the full dataset (11G) I found the "loco-v3-darknet-85c.zip". In that "975902,7266.txt" and others are contained, are these the final annotations and can be used?

Hi,

I have compiled the dataset to train with Yolov8. It took me some time to run through the current data hierarchy and compile all images and annotations at the same place and I was thinking to contribute them here to save people's efforts and time who are trying this out with Yolo family of models.

I have compiled the data in the following Yolo structure:

data
|__images
|train
|<train_ids>.jpg
|test
|<test_ids>.jpg

|__labels
|train
|<train_labels>.txt
|test
|<test_labels>.txt

Can you please suggest how to contribute the dataset here?

Hi there,

I am currently working on this dataset in the context of my thesis. While reading the paper, I was wondering if the video that is discussed in section IV C "Preliminary Analysis" and referenced to be in this repo is available? I might be new to GitHub, and maybe I am overseeing it, so a little help would be much appreciated.