Common-Objects-in-Hemispherical-Images-Dataset

Common Objects in Hemispherical Images (COHI) is a benchmark testing dataset for object detection in hemispherical/fisheye cameras. It contains 1,000 real fisheye images of 39 classes sampled from the MS COCO dataset with 14.2 object instances per image. The images were captured using a hemispherical camera ELP-USB8MP02G-L180 with the 2,448 by 3,264 pixel resolution and manually annotated with standard axis-aligned bounding boxes afterward. The samples of raw and annotated images from the COHI dataset are shown below.

The names of sampled classes and the number of bounding boxes for each class are presented in the next figure.

The FisheyeCOCO dataset

To train object detection models for the COHI dataset, the FisheyeCOCO dataset was generated by applying a non-linear mapping to the MS COCO dataset to obtain fisheye-looking images. The samples of raw and annotated images from the FisheyeCOCO dataset are illustrated below.

Download the datasets

The COHI dataset

One can access the COHI dataset using the following link: COHI.

The FisheyeCOCO dataset

The FisheyeCOCO dataset can be downloaded with a link: FisheyeCOCO. Alternatively, it can be generated from the MS COCO 2017 dataset using the scripts fisheye_transform.py and coordinates_transform.py in the preprocessing directory.

Requirements

For data preprocessing

• numpy
• PIL
• pandas

For object detection

We used YOLOv7 to train and evaluate object detection models. All needed information can be found on their official GitHub page YOLOv7.

Pre-trained models

We trained the YOLOv7 model with 36.9 M parameters on three datasets and evaluated the performance of models with our benchmark testing dataset - COHI.

• YOLOv7_original: trained on the MS COCO dataset
• YOLOv7_transformed: trained on the FisheyeCOCO dataset
• YOLOv7_combined: trained on the combination of the MS COCO and FisheyeCOCO datasets

mAP₅₀ results are summarized in the table below.

Model	mAP50
YOLOv7_original	58.23%
YOLOv7_transformed	58.28%
YOLOv7_combined	60.77%

Pre-trained model weights can be downloaded using a Google Drive link: weights.

Citation

@article{Balgabekova2023,
author = "Zarema Balgabekova and Muslim Alaran and Hüseyin Atakan Varol",
title = "{A Data-Centric Approach for Object Recognition in Hemispherical Camera Images}",
year = "2023",
month = "5",
url = "https://www.techrxiv.org/articles/preprint/A_Data-Centric_Approach_for_Object_Recognition_in_Hemispherical_Camera_Images/23016185",
doi = "10.36227/techrxiv.23016185.v1"
}