ParkingSticker: A Real-World Object Detection Dataset

TL;DR

ParkingSticker is a challenging real-world object detection dataset with small objects, derived from security camera footage, designed to test and improve industry-relevant detection methods.

Contribution

The paper introduces ParkingSticker, a new dataset that closely mimics real-world industry scenarios with small objects and limited data, filling a gap in existing datasets.

Findings

YOLOv2 achieves feasible detection performance

Small object detection remains challenging in real-world data

Dataset encourages development of robust detection methods

Abstract

We present a new and challenging object detection dataset, ParkingSticker, which mimics the type of data available in industry problems more closely than popular existing datasets like PASCAL VOC. ParkingSticker contains 1,871 images that come from a security camera's video footage. The objective is to identify parking stickers on cars approaching a gate that the security camera faces. Bounding boxes are drawn around parking stickers in the images. The parking stickers are much smaller on average than the objects in other popular object detection datasets; this makes ParkingSticker a challenging test for object detection methods. This dataset also very realistically represents the data available in many industry problems where a customer presents a few video frames and asks for a solution to a very difficult problem. Performance of various object detection pipelines using a YOLOv2 architecture are presented and indicate that identifying the parking stickers in ParkingSticker is challenging yet feasible. We believe that this dataset will challenge researchers to solve a real-world problem with real-world constraints such as non-ideal camera positioning and small object-size-to-image-size ratios.