A synthetic texture-invariant dataset for object detection of UAVs

Last update: Aug 13, 2022

Overview

A synthetic dataset for object detection of UAVs

This repository contains a synthetic datasets accompanying the paper Sim2Air - Synthetic aerial dataset for UAV monitoring by Antonella Barisic, Frano Petric and Stjepan Bogdan.

In this paper, we propose to use a texture-invariant representation of objects for aerial object detection. Our approach improves the generalisation and robustness of the object detector. A dataset is created with randomly assigned atypical textures and sufficient diversity and photorealism in all other components such as shape, pose, lighting, scale, background, etc. The results also show improved accuracy in case of distant objects and difficult lighting conditions.

All datasets from the paper are available for download. If you use these datasets for your research, please cite:

@misc{barisic2021sim2air,
      title={Sim2Air - Synthetic aerial dataset for UAV monitoring}, 
      author={Antonella Barisic and Frano Petric and Stjepan Bogdan},
      year={2021},
      eprint={2110.05145},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Datasets

Name	Description
Synthetic Eagle Baseline (SEB)	The SEB dataset is a synthetic dataset with a single UAV model, the custom aerial platform Eagle. Since this dataset serves as the basis for proving our hypothesis, it was created with only one texture, identical to the texture of real-life Eagle. SEB consists of 32 000 images of size 604 x 604 with annotations in YOLO format.
Synthetic Eagle with Textures (SET)	The SET dataset is the main star ⭐ of our work. It is a synthetic dataset of a single model, the custom aerial platform Eagle, with randomly selected atypical textures. The mixture of 32 different textures is applied during the procedural generation of the dataset. SET also consists of 32 000 images of size 604 x 604 with annotations in YOLO format.
Synthetic UAVs with Textures (S-UAV-T)	The S-UAV-T dataset is similar to SET but with many more models of UAVs. The data was created with 10 different multicopter models, 32 atypical textures, and with a variety of poses, backgrounds, viewpoints, etc. S-UAV-T consists of 52 500 images of size 604 x 604 with annotations in YOLO format.

If you want to test your detection results against real data, check out our UAV-Eagle dataset at larics/UAV-Eagle.

Contact

For more information, please contact Antonella Barisic.

A synthetic texture-invariant dataset for object detection of UAVs

Related tags

Overview

A synthetic dataset for object detection of UAVs

Datasets

Contact

Owner

LARICS Lab

We evaluate our method on different datasets (including ShapeNet, CUB-200-2011, and Pascal3D+) and achieve state-of-the-art results, outperforming all the other supervised and unsupervised methods and 3D representations, all in terms of performance, accuracy, and training time.

traiNNer is an open source image and video restoration (super-resolution, denoising, deblurring and others) and image to image translation toolbox based on PyTorch.

A Streamlit demo demonstrating the Deep Dream technique. Adapted from the TensorFlow Deep Dream tutorial.

Cosine Annealing With Warmup

PyTorch Implementation of Fully Convolutional Networks. (Training code to reproduce the original result is available.)

Tutorials, assignments, and competitions for MIT Deep Learning related courses.

Code for the paper "Implicit Representations of Meaning in Neural Language Models"

Chainer Implementation of Semantic Segmentation using Adversarial Networks

This repo is customed for VisDrone.

Pacman-AI - AI project designed by UC Berkeley. Designed reflex and minimax agents for the game Pacman.

New approach to benchmark VQA models

Code for paper "A Critical Assessment of State-of-the-Art in Entity Alignment" (https://arxiv.org/abs/2010.16314)

Self-Regulated Learning for Egocentric Video Activity Anticipation

A denoising diffusion probabilistic model synthesises galaxies that are qualitatively and physically indistinguishable from the real thing.

Codes for realizing theories learned from Data Mining, Machine Learning, Deep Learning without using the present Python packages.

Continuous Diffusion Graph Neural Network

A robust pointcloud registration pipeline based on correlation.

Energy consumption estimation utilities for Jetson-based platforms

Source code of "Hold me tight! Influence of discriminative features on deep network boundaries"

Flower - A Friendly Federated Learning Framework