LaB-CL: Localized and Balanced Contrastive Learning for improving parking slot detection

TL;DR

This work proposes the first supervised contrastive learning framework for parking slot detection, Localized and Balanced Contrastive Learning for improving parking slot detection (LaB-CL), and proposes a new hard negative sampling scheme that selects local representations with high prediction error.

Abstract

Parking slot detection is an essential technology in autonomous parking systems. In general, the classification problem of parking slot detection consists of two tasks, a task determining whether localized candidates are junctions of parking slots or not, and the other that identifies a shape of detected junctions. Both classification tasks can easily face biased learning toward the majority class, degrading classification performances. Yet, the data imbalance issue has been overlooked in parking slot detection. We propose the first supervised contrastive learning framework for parking slot detection, Localized and Balanced Contrastive Learning for improving parking slot detection (LaB-CL). The proposed LaB-CL framework uses two main approaches. First, we propose to include class prototypes to consider representations from all classes in every mini batch, from the local perspective. Second, we propose a new hard negative sampling scheme that selects local representations with high prediction error. Experiments with the benchmark dataset demonstrate that the proposed LaB-CL framework can outperform existing parking slot detection methods.

Figures (6)

• Figure 1: Illustration of the proposed LaB-CL framework for parking slot detection. The $\bullet$ denotes local representations and the $\bigstar$ denotes a local class prototype for each class, and different colors mean that that instances and class prototypes are from different classes. Proposed LaB-CL moves all the within-class local representations toward their local prototypes that are equidistant between each other while treating all classes equally.
• Figure 2: Examples of junctions in parking slots. The black dot denotes the junction point in parking slots that is parameterized by its $x$- and $y$-coordinates. The red dotted line denotes the rotational position of each junction that is parameterized with an angle $\theta$ from the zero angle in radians.
• Figure 3: Illustration of the global and local representations. (a) The global representation is extracted from the entire image. (b) The local representations extracted from the entire image correspond to patches in the original image space.
• Figure 4: Overview of the proposed LaB-CL framework for junction detection of parking slots. The proposed framework simultaneously learns local representations and detector in an E2E learning manner. (Localized CL) We propose to use local representations rather than global representations; see Section \ref{['subsec:Preliminaries']}. (Balanced CL) We use transformed classifier weights as localized class prototypes. In addition, we use memory banks with the proposed hard negative sampling strategy. We include these samples in every minibatch to consider samples from all classes, i.e., balanced CL. (Postprocessing for parking slot detection) Some postprocessing schemes are applied to detected junctions to identify pairs of junctions and detect parking slots DMPR-PS.
• Figure 5: Qualitative parking slot detection results under various conditions. Red indicates the junction, and blue indicates the entrance and side lines.