Extended Agriculture-Vision: An Extension of a Large Aerial Image Dataset for Agricultural Pattern Analysis

Extended Agriculture-Vision (Published in TMLR):

A continuous work of Agriculture-Vision, with great collaborators to bring Agriculture and Computer Vision / AI communities together to benefit humanity!

Overview

1. Paper
2. Dataset
3. Quick Start

1. Papers

Extended Agriculture-Vision on TMLR:

TMLR

ArXiv

Supplementary

@article{
wu2023extended,
title={Extended Agriculture-Vision: An Extension of a Large Aerial Image Dataset for Agricultural Pattern Analysis},
author={Jing Wu and David Pichler and Daniel Marley and Naira Hovakimyan and David A Wilson and Jennifer Hobbs},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2023},
url={https://openreview.net/forum?id=v5jwDLqfQo},
note={}
}

@article{wu2023extended,
title={Extended Agriculture-Vision: An Extension of a Large Aerial Image Dataset for Agricultural Pattern Analysis},
author={Wu, Jing and Pichler, David and Marley, Daniel and Wilson, David and Hovakimyan, Naira and Hobbs, Jennifer},
journal={arXiv preprint arXiv:2303.02460},
year={2023}
}

2. Dataset

The Extended Agriculture-Vision can be downloaded with the following command:

aws s3 cp  s3://intelinair-data-releases/agriculture-vision/TMLR_benchmarks_2023/ Path_to_Save --no-sign-request --recursive

For example, to download the dataset in your current location, use:

aws s3 cp  s3://intelinair-data-releases/agriculture-vision/TMLR_benchmarks_2023/ . --no-sign-request --recursive

Previous supervised dataset can be downloaded with the following command:

aws s3 cp s3://intelinair-data-releases/agriculture-vision/cvpr_challenge_2021/supervised supervised --no-sign-request --recursive

3. Quick Start on Pre-trained Weights

3.1 Pre-trained Models

Dataset	Pre-trained Methods	Architecture	Link
Extended AgVision	MoCoV2	ResNet-18	download
Extended AgVision	MoCoV2	ResNet-50	download

3.2 To load pre-trained ResNet-18 with RGBN channels (The same when applied to ResNet-50,101)

import torch
import torchvision.transforms as transforms
import torchvision.models as models

# Create a new ResNet-18 model with four channels input
resnet18_four_channels = models.resnet18(pretrained=False)
resnet18_four_channels.conv1 = torch.nn.Conv2d(4, 64, kernel_size=7, stride=2, padding=3, bias=False)
resnet18_four_channels = torch.nn.Sequential(*list(resnet18_four_channels.children())[:-1], nn.Flatten())

# Load the saved weights into the new model
resnet18_four_channels.load_state_dict(torch.load('Res_18.pth'))

3.3 To load pre-trained ResNet-18 with RGB channels (The same when applied to ResNet-50,101)

import torch
import torchvision.transforms as transforms
import torchvision.models as models

# Create a new ResNet-18 model with three channels input
resnet18_three_channels = models.resnet18(pretrained=False)
resnet18_three_channels = torch.nn.Sequential(*list(resnet18_three_channels.children())[:-1], nn.Flatten())

# Load the saved weights from the model with four channels
saved_state_dict = torch.load('Res_18.pth')
# Discard the extra channel in the weights (assuming the first channel needs to be discarded)
saved_state_dict['0.weight'] = saved_state_dict['0.weight'][:, :-1, :, :]

# Load the modified state_dict into the new model
resnet18_three_channels.load_state_dict(saved_state_dict)