AI4Agriculture Counting Model
A computer vision model to automatically detect grape bunches in vineyard pictures.
The AI4Agriculture Counting Model is a computer vision model to automatically detect grape bunches in vineyard pictures. The bunches are detected by using bounding boxes. This model can be used in viticulture applications, such as yield estimation, fruit counting, and field robotics.
The proposed solution is based on a Cascade-RCNN convolutional neural network (from the MMDetection Toolbox) trained on the Ai4Agriculture dataset.
Hardware Architecture:
x64 with CUDA support (optional, preferred)
Install and Run:
To download the docker container use the following command line:
docker pull upcai4eu/ai4counting:latest
To test it, run it with one of the following commands:
docker run --gpus all -p 8061:8061 -it upcai4eu/ai4counting
If the hardware used for the tests does not provide CUDA support use instead this command:
docker run -p 8061:8061 -it upcai4eu/ai4counting
Detailed description, Protocol Buffers interfaces description and updated technical data can be found in the AI4EU-AgriCounting Model at the AI4EU Experiments platform.
User guide and sample client
AI4EU-AgriCounting Model
Additional information:
A Cascade-RCNN model, pretrained on the COCO 2017 database has been fine-tuned using the Ai4Agriculture dataset. The model uses a ResNet50 backbone. The Cascade-RCNN is a mulkti-stage variant of the Faster-RCNN model. The stages are trained sequentially using increasing Intersection over Union (IoU) thresholds. This strategy provides more selectivity against close false positives. At inference time, the same cascaded architecture is used.
The MMDetection implementation of the Cascade RCNN has been used. MMDetection is an open source object detection toolbox based on PyTorch. It is a part of the OpenMMLab project.