Comparision of YOLOV9 instance segmentation and SAM based segmentation on remote sensing images

YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information

Web application

Pull Docker image

docker pull krishnatejan/yolov9_instance_segmentation:latest

Run container

docker run -p 8080:5000 --name instance --gpus all krishnatejan/yolov9_instance_segmentation 

YoloV9 Object detection + SAM Instance segmentation

yolov9-c.pt - For training the object detection task by using transfer learning .

python train_dual.py --workers 8 --device 0 --batch 2 --data data/dataset.yaml --img 1024 --cfg models/detect/yolov9-c.yaml --name yolov9-c --hyp hyp.scratch-high.yaml --epochs 20 --optimizer Adam --weights ./yolov9-c.pt 

Segment Anything Model (SAM) is used to segment the image based on boundary boxes provided by the yolo9-c model. Follow instructions to download the model from here

python detect_dual.py --source './data/images/bike.jpg' --img 640 --device 0 --weights './yolov9-c.pt' --name yolov9_c_640_detect 

Instance Segmentation using GELAN (YoLoV9)

Training

Experiment 1: Images are super resolved using RealESRGAN and also annotations are scaled.

python segment/train.py --device 0 --batch 2  --data dataset.yaml --img 1024 --cfg models/segment/gelan-c-seg.yaml --name gelan-c-seg --hyp hyp.scratch-high.yaml --no-overlap --epochs 50 --optimizer Adam --workers 8

Experiment 2: Images are kept the same size and trained

python segment/train.py --device 0 --batch 4  --data dataset.yaml --img 512 --cfg models/segment/gelan-c-seg.yaml --name gelan-c-seg-512 --hyp hyp.scratch-high.yaml --no-overlap --epochs 20 --optimizer Adam --workers 8

Experiment 3: Mosaic Augumentation wa applied on dataset with original dataset to observe the effect

python segment/train.py --device 0 --batch 2  --data dataset.yaml --img 1024 --cfg models/segment/gelan-c-seg.yaml --name gelan-c-seg-aug --hyp hyp.scratch-high.yaml --no-overlap --epochs 20 --optimizer Adam --workers 8

Inference

python ./segment/val.py --data data/dataset.yaml --img 1024 --batch 8 --conf 0.001 --iou 0.7 --device 0 --weights './"E:/yolov9/runs/train-seg/gelan-c-seg-1024/weights/best.pt"' --save-json --name gelan_c_seg_1024_val --verbose

python ./segment/val.py --data data/dataset.yaml --img 1024 --batch 8 --conf 0.001 --iou 0.7 --device 0 --weights './"E:/yolov9/runs/train-seg/gelan-c-seg-aug/weights/best.pt"' --save-json --name gelan_c_seg_aug_val --verbose

python ./segment/val.py --data data/dataset.yaml --img 512 --batch 8 --conf 0.001 --iou 0.7 --device 0 --weights './"E:/yolov9/runs/train-seg/gelan-c-seg-512/weights/best.pt"' --save-json --name gelan_c_seg_512_val --verbose

Predict

python ./segment/predict.py --data data/dataset.yaml --img 512 --conf 0.001 --iou 0.7 --device 0 --weights './"E:/yolov9/runs/train-seg/gelan-c-seg-512/weights/best.pt"' --source "E:\yolov9\dataset\images\test\thueringen_DETHL54P0000w17m_0_4c5915a7-573a-4215-b65c-cea3764d7837.image_dop_10_400000.png" --hide-labels --max-detect 15

Note

These models are only trined for 20 epochs because of lack of computational power hence the accuracy of results.