FoveaBox (IxRT)

Model Description

FoveaBox is an advanced anchor-free object detection framework that enhances accuracy and flexibility by directly predicting the existence and bounding box coordinates of objects. Utilizing a Feature Pyramid Network (FPN), it adeptly handles targets of varying scales, particularly excelling with objects of arbitrary aspect ratios. FoveaBox also demonstrates robustness against image deformations.

Supported Environments

Model Preparation

Prepare Resources

Pretrained model: https://download.openmmlab.com/mmdetection/v2.0/foveabox/fovea_r50_fpn_4x4_1x_coco/fovea_r50_fpn_4x4_1x_coco_20200219-ee4d5303.pth

Dataset: http://images.cocodataset.org/zips/val2017.zip to download the validation dataset.

Install Dependencies

# Install libGL
## CentOS
yum install -y mesa-libGL
## Ubuntu
apt install -y libgl1-mesa-glx

pip3 install -r requirements.txt

Model Conversion

# export onnx model
python3 export.py --weight fovea_r50_fpn_4x4_1x_coco_20200219-ee4d5303.pth --cfg fovea_r50_fpn_4xb4-1x_coco.py --output foveabox.onnx

# Use onnxsim optimize onnx model
onnxsim foveabox.onnx foveabox_opt.onnx

Model Inference

export DATASETS_DIR=/Path/to/coco/

FP16

# Accuracy
bash scripts/infer_foveabox_fp16_accuracy.sh
# Performance
bash scripts/infer_foveabox_fp16_performance.sh

Model Results

References

• mmdetection