RLB Object Detection

Based on FastestDet by dog-qiuqiu.

Requirements

Run install.sh to create a Conda environment and to install other dependencies.

Dataset

Formatting

Dataset should be formatted like any other YOLO dataset:

.
├── train
│   ├── 000001.jpg
│   ├── 000001.txt
│   ├── 000002.jpg
│   ├── 000002.txt
│   ├── 000003.jpg
│   └── 000003.txt
└── val
    ├── 000043.jpg
    ├── 000043.txt
    ├── 000057.jpg
    ├── 000057.txt
    ├── 000070.jpg
    └── 000070.txt

Annotation files can contain multiple detections, one detection per line following the class_id x y w h format.

Example of a single annotation file:

11 0.344192634561 0.611 0.416430594901 0.262
14 0.509915014164 0.51 0.974504249292 0.972

RLB dataset

• Download RLB dataset
• Extract the downloaded archive
• Run the build.sh script to generate train and test folders
• Move these folders into a RLB_dataset folder
• Move the RLB_dataset folder into the dataset folder of this repo

Note

This dataset contains augmented data using this dataset as background images.

COCO2017 dataset

• Download COCO 2017 train and val images/annotations from here
• Use JSON2YOLO to convert instances_train2017.json and instances_val2017.json to YOLO format
• Extract train images into a train folder, do the same for val images into a test folder
• Move these folders into a COCO2017_dataset folder
• Move the COCO2017_dataset folder into the dataset folder of this repo

Optimizations

Use onnx_to_int8_ncnn.sh.

Example for a model.onnx file:

./onnx_to_int8_ncnn.sh ./checkpoints/2024-06-04_16-18-39/20_0.24095053259538232