CVPR 2021 Challenge on Agriculture Vision

This repo contains the code to reproduce our results in CVPR21 Challenge on Agriculture-Vision. We ranked 4th in the supervised track.

By Songyao Jiang, Bin Sun, and Yun Fu, from Smile Lab @ Northeastern University

Introduction

The first model is modified MSCG-Net, please see MSCG-Net/README.md to train and test the model. The second model is modified DeepLabv3, please see Deeplabv3_Ensemble/Readme.txt to train and test the model. The results of the above models are assembled together to improve the overall mIoU using the ensemble code in Deeplabv3_Ensemble. We used ensemble results as our final submission during the challenge

Pretrained models

Google Drive

Code structure

├── MSCGNet                # Model 1
├── Deeplabv3_Ensemble	   # Model 2 and ensemble
└── challenge_report       # Detailed report submitted

Results Summary

Model	Backbone	#Params	mIoU
MSCG-Net	ResNet-101	31M	0.464
DeepLabv3	ResNet-101	60M	0.494
Ensemble	N/A	91M	0.507

Summary of AgriVision dataset 2021

Splits: 56,944/18,334/19,708 train/val/test

Resolution: 512 x 512

Modalities: 1. RGB, 2. NIR (Near-infrared)

Annotations:

0 - background, 1 - double_plant, 2 - drydown, 3 - endrow, 4 - nutrient_deficiency, 5 - planter_skip, 6 - water, 7 - waterway, 8 - weed_cluster

---

MSCG-Net for Semantic Segmentation

This model is modified from MSCG-Net models (MSCG-Net-50 and MSCG-Net-101) for semantic segmentation in Agriculture-Vision Challenge and Workshop (CVPR 2021).

Pretrained model

https://drive.google.com/file/d/1oW503NxUfwANfKQZ8zT3gG_XDWSuwwsQ/view?usp=sharing

DeepLabv3 with Ensemble

This folder contains code modified from Deeplabv3 for the CVPR 2021 Challenge on Agriculture Vision. This folder also contains ensemble code to obtain our final results.

Pretrained models

https://drive.google.com/drive/folders/1VnPKVErUHEjbCe5ailsSvhXCjZWnw0qH?usp=sharing

Reference

https://github.com/samleoqh/MSCG-Net

https://github.com/LAOS-Y/AgriVision

https://github.com/HRNet/HRNet-Semantic-Segmentation