Inception Robotics SAR Object Detection Challenge

00:00:00

Overview

The objective of this challenge is to achieve superior mAP@50 scores on the SARDet-100K dataset, surpassing the performance demonstrated in the SARDet-100K paper

Description

Synthetic Aperture Radar (SAR) images are an integral part of remote sensing, well-suited for capturing geographical images under any weather condition, land cover, camouflage, etc. Detecting objects in SAR images is challenging due to the limited availability of data, low resolution, small size of objects, and the significant domain gap when adapting models pre-trained on natural RGB datasets. These challenges are explained in detail in the paper.

For this challenge, participants are expected to build on the improvements in mAP demonstrated in the SARDet-100k paper. Many key insights on handling SAR images are already provided in the paper. Participants must leverage these insights (e.g., using handcrafted features using the Wavelet Scattering Transform) and bring in their own ideas to train an object detection model that establishes a new state-of-the-art performance on the SARDet-100k dataset. Models need not be limited to compact CNN or Transformer-based models. Participants are encouraged to also leverage the latest open-source large multimodal models (e.g., GLIP, Grounding DINO, Llama3, Florence).


Dataset

The GitHub repository for the dataset and a SAR object detection model can be found here

Deliverable

Participants must submit the following deliverables by 22nd July 2024 11:59 PM ET

Rewards

For teams that surpass the mAP@50 scores in the SARdet-100K paper (mAP@50 of 84.0)

NOTES

Registration

Contact Us

For any queries, please contact us at Inception Robotics

Rules