This dataset contains image data derived from the Objects365 object-detection dataset and reorganized for an 80-class detection/pretraining pipeline.

The original Objects365 dataset was introduced by Shao et al. at ICCV 2019 as a large-scale object-detection dataset containing 365 object categories and more than 10 million manually annotated bounding boxes [1].

Important: This repository/dataset is a derived 80-class organization of Objects365 data. It is not the official Objects365 distribution, and the directory structure and JSONL metadata described below are specific to this dataset organization.

For information about the original dataset, see the official Objects365 project and publication [1, 2].

Image Data Organization

This document describes the physical organization of the image data used by this dataset.

The image storage is intentionally separated from the label metadata. Labels are stored as JSONL files, while the actual image files are organized into split-specific directories and smaller patch subdirectories.

The patch structure described below is a storage convention used by this dataset and should not be interpreted as part of the official Objects365 annotation specification.

1. Directory Structure

The image dataset is divided into two main splits:

train/ — training images
val/ — validation images

Each split is further divided into multiple patch directories. Each patch contains a subset of the images belonging to that split.

images/
├── train/
│   ├── patch0/
│   │   ├── image_a.jpg
│   │   ├── image_b.jpg
│   │   └── ...
│   ├── patch1/
│   │   ├── image_c.jpg
│   │   └── ...
│   ├── patch2/
│   │   └── ...
│   └── ...
└── val/
    ├── patch0/
    │   ├── image_x.jpg
    │   └── ...
    ├── patch1/
    │   └── ...
    ├── patch2/
    │   └── ...
    └── ...

The exact number of patch directories may vary. The loader should therefore rely on the path metadata rather than assuming a fixed patch range.

2. Why Images Are Split into Patches

Instead of placing every image from a split into one very large directory, images are distributed across multiple patch directories.

This organization provides several practical advantages:

avoids extremely large single directories;
makes file management and transfer easier;
keeps the physical dataset structure manageable at large scale;
allows the loader to locate an image directly from its recorded relative path.

The patch directory itself has no semantic meaning. It is only a physical storage partition.

3. Physical Image Path Metadata

The label metadata contains one JSONL file that maps each image filename to its physical location inside the corresponding image split.

For training data, the mapping is stored in:

labels/train/images_train.jsonl

For validation data:

labels/val/images_val.jsonl

Each line contains an independent JSON object:

{"image_name": "objects365_v2_00953995.jpg", "path": "patch17/objects365_v2_00953995.jpg"}

The path field is a relative path.

It is interpreted relative to the root directory of the corresponding image split.

For example:

image split root:
images/train/

relative path:
patch17/objects365_v2_00953995.jpg

The final physical path is:

images/train/patch17/objects365_v2_00953995.jpg

Likewise, for validation:

image split root:
images/val/

relative path:
patch3/objects365_v1_00123456.jpg

becomes:

images/val/patch3/objects365_v1_00123456.jpg

Therefore, the JSONL mapping does not store an absolute filesystem path and does not include the train/ or val/ prefix.

4. Relationship Between Image Metadata and Physical Files

The dataset links logical image metadata to physical image files through the filename.

The relationship is:

images_info.jsonl
    file_name
        │
        │ exact filename match
        ▼
images_train.jsonl / images_val.jsonl
    image_name
        │
        ▼
    relative path
        │
        ▼
train/patchXX/... or val/patchXX/...

Example:

// images_info.jsonl
{
  "id": 900001,
  "file_name": "objects365_v2_00900001.jpg",
  "width": 333,
  "height": 500
}

// images_train.jsonl
{
  "image_name": "objects365_v2_00900001.jpg",
  "path": "patch12/objects365_v2_00900001.jpg"
}

The loader can then resolve:

images/train/patch12/objects365_v2_00900001.jpg

This separation keeps logical metadata independent from the physical location of the image on disk.

5. Path Resolution in the DataLoader

At runtime, the loader performs the following steps:

image_id
   │
   ▼
images_info.jsonl
   │
   └── file_name
          │
          ▼
images_train.jsonl / images_val.jsonl
          │
          └── relative path
                 │
                 ▼
split root + relative path
                 │
                 ▼
physical image file

Conceptually:

rel_path = image_path_map[file_name]

image_path = os.path.join(
    images_root_dir,
    images_split_dir,
    rel_path,
)

For the training split:

images_split_dir = "train"

For the validation split:

images_split_dir = "val"

6. Design Principle

The dataset deliberately separates three concerns:

Logical image metadata
    └── images_info.jsonl

Object annotations
    └── annotations.jsonl

Physical image location
    └── images_train.jsonl / images_val.jsonl

The physical image directory only stores image files:

train/patchXX/*.jpg
val/patchXX/*.jpg

All relationships between image IDs, filenames, labels, and physical paths are maintained through the JSONL metadata.

This structure makes the dataset easier to reorganize or relocate because the physical storage layout can change independently from the annotation schema, as long as the image path mapping is updated accordingly.

Source Dataset

The image data originates from the Objects365 dataset.

Objects365 was introduced by Shao et al. as a large-scale benchmark and pretraining dataset for object detection. The original ICCV 2019 release contains 365 object categories, approximately 638,000 images, and more than 10 million bounding-box annotations [1].

The official Objects365 project maintains information about subsequent dataset releases, downloads, licensing, and dataset statistics [2, 3].

This 80-class dataset is a derived subset/reorganization and should therefore be cited together with the original Objects365 publication.

License and Attribution

The licensing terms of the original Objects365 dataset must be respected when using or redistributing derived data.

According to the official Objects365 download page [3]:

Objects365 annotations and the Objects365 website are licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
The Objects365 Consortium does not own the copyright of the underlying images.
Image usage remains subject to the applicable original image/Flickr terms and copyright restrictions.
Users are responsible for ensuring that their use of the images complies with the relevant copyright and licensing requirements.
The official Objects365 terms state restrictions concerning redistribution of the original images.

Therefore, inclusion of metadata or transformed annotations in this dataset does not transfer ownership or copyright of the source images.

Users intending to redistribute, publish, commercially use, or otherwise repurpose the image data should independently verify that their intended use complies with the original Objects365 terms and the rights associated with the individual images.

See the official Objects365 license/download page for the authoritative licensing terms [3].

Citation

If this dataset is used in research, please cite the original Objects365 publication:

@inproceedings{shao2019objects365,
  author    = {Shuai Shao and
               Zeming Li and
               Tianyuan Zhang and
               Chao Peng and
               Gang Yu and
               Xiangyu Zhang and
               Jing Li and
               Jian Sun},
  title     = {Objects365: A Large-Scale, High-Quality Dataset for Object Detection},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2019},
  pages     = {8430--8439}
}

If results specifically depend on the 80-class filtering, metadata conversion, JSONL representation, or patch-based image layout provided by this dataset, the derived dataset should also be identified separately in the experimental setup so that the preprocessing pipeline can be reproduced.

References

[1] S. Shao, Z. Li, T. Zhang, C. Peng, G. Yu, X. Zhang, J. Li, and J. Sun,
“Objects365: A Large-Scale, High-Quality Dataset for Object Detection,”
Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 8430–8439.
https://openaccess.thecvf.com/content_ICCV_2019/html/Shao_Objects365_A_Large-Scale_High-Quality_Dataset_for_Object_Detection_ICCV_2019_paper.html

[2] Objects365 Consortium,
Objects365 Dataset — Official Website.
https://www.objects365.org/

[3] Objects365 Consortium,
Objects365 Dataset — Download and License Information.
https://www.objects365.org/download.html

Downloads last month: 22

Total file size:

377 GB