A Self-Supervised Approach for Enhanced Feature Representations in Object Detection Tasks

TL;DR

This work presents a model trained on unlabeled data that outperforms state-of-the-art feature extractors pre-trained on ImageNet and particularly designed for object detection tasks, and encourages the model to focus on the most relevant aspects of an object, thus achieving better feature representations and, therefore, reinforcing its reliability and robustness.

Abstract

In the fast-evolving field of artificial intelligence, where models are increasingly growing in complexity and size, the availability of labeled data for training deep learning models has become a significant challenge. Addressing complex problems like object detection demands considerable time and resources for data labeling to achieve meaningful results. For companies developing such applications, this entails extensive investment in highly skilled personnel or costly outsourcing. This research work aims to demonstrate that enhancing feature extractors can substantially alleviate this challenge, enabling models to learn more effective representations with less labeled data. Utilizing a self-supervised learning strategy, we present a model trained on unlabeled data that outperforms state-of-the-art feature extractors pre-trained on ImageNet and particularly designed for object detection tasks. Moreover, the results demonstrate that our approach encourages the model to focus on the most relevant aspects of an object, thus achieving better feature representations and, therefore, reinforcing its reliability and robustness.

Figures (5)

• Figure 1: Simplification of the SimCLR algorithm. For the sake of clarity, we have represented the model multiple times, each applied to a different input image.
• Figure 2: Top-1 and Top-3 classification accuracies for both datasets.
• Figure 3: Localization accuracies at IoU 0.5 and 0.7 for both datasets.
• Figure 4: Differences between the Baseline and the SSL backbone over the number of images per class for both datasets.
• Figure 5: Heat maps generated by Grad-CAM for five representative samples from the PascalVOC datasets. The top row shows the activations from the Baseline model, highlighting its focus on specific, fragmented regions of the object. The bottom row illustrates the activations of the SSL backbone, capturing the entire object shape and its defining characteristics.