Journal Article

OBJECT-CENTRIC MASKED IMAGE MODELLING-BASED SELF-SUPERVISED PRE TRAINING FOR REMOTE SENSING OBJECTS DETECTION

Nagagopiraju, V, Reddy, P Manohar, Ramadas, K Venkata, Rajini, M, Pullarao, B

2025

Overview

Masked Image Modelling (MIM) is an effective self-supervised pre-training (SSP) task that helps models learn task-agnostic representations, improving fine-tuning for downstream tasks. However, in remote sensing, high random masking ratios hinder the capture of small objects and local details, causing misalignment between pre-training and fine-tuning. To address this, we propose Object-Centric MIM (OCMIM) for better object-level representation learning. First, an object-centric data generator automatically sets up targeted pre-training data based on object characteristics, optimizing object detection. Second, an attention-guided mask generator enhances MIM by focusing on highly attended object regions instead of random masking. This approach improves feature discrimination and object detection performance. Experiments on six remote sensing benchmarks demonstrate that OCMIM-based SSP outperforms traditional methods, making it a superior pre training strategy for remote sensing object detection.

• Publisher: Kohat University of Science and Technology (KUST)
• Subject: Accuracy / Datasets / Efficiency / Masking / Misalignment / Modelling / Neural networks / Object recognition / Optimization / Remote sensing / Representation learning / Telematics / Unmanned aerial vehicles