Awesome

LFG-Net: Low-level Feature Guided Network for Precise Ship Instance Segmentation in SAR Images

This project hosts the code for reproducing experiment results of LFG-Net

LFG-Net is based on mmdetection framework. Please follow the official guideline of installing the prerequisites

Highlights

Small object instance segmentation framework for SAR images.
Enhancing the low-level features from image level to instance level.
LFG-Net achieves state-of-the-art instance segmentation performance on HRSID, SSDD, and AirSARShip dataset.

Requirements

ubuntu == 18.04
mmdetection == 2.20
mmcv == 1.4.2
torch-dct
pytorch == 1.7.0, torchvision == 0.8.1

Usage

Training

To train LFG-Net model with original settings of our paper, run:

python train.py

Inference

To inference the trained model with a single gpu, run:

python tools/test.py <CONFIG_FILE> <DET_CHECKPOINT_FILE> --eval bbox segm

Performance on HRSID

Ship instance segmentation results on HRSID. The models are trained for 12 epochs with the initial learning rate at 0.0025. Results are evaluated with MS COCO evaluation metrics, Parameters, and FPS on Quadaro RTX 6000.

Model	AP	AP<sub>50</sub>	AP<sub>75</sub>	AP<sub>S</sub>	AP<sub>M</sub>	AP<sub>L</sub>	Params.	FPS
SOLO	13.8	27.8	13.8	14.2	13.6	3.8	54.92M	14.2
Yolact	35.3	67.2	35.2	34.1	50.0	6.2	34.73M	17.3
Mask R-CNN	52.2	80.6	63.7	51.8	61.3	9.9	43.75M	14.9
Point Rend	53.8	81.8	65.3	53.1	63.3	15.5	55.53M	12.3
GRoIE	52.0	79.7	62.8	51.4	61.1	17.5	47.54M	8.3
Mask Scoring R-CNN	53.0	80.8	63.8	52.6	61.0	11.8	60.01M	14.7
R-ARE-Net	53.6	80.4	65.9	55.3	55.2	13.5	46.58M	10.4
QueryInst	44.2	69.5	53.2	43.4	54.6	12.2	172.22M	4.5
Cascade Mask R-CNN	53.3	82.0	64.0	52.7	61.9	18.3	76.08M	13.5
Hybrid Task Cascade	53.6	82.3	64.7	52.8	63.2	18.6	79.73M	9.7
Detectors	54.1	82.4	65.5	53.3	64.2	20.7	134.00M	6.3
SCNet	54.4	82.4	65.9	54.1	62.1	13.2	94.29M	8.6
LFG-Net	59.7	88.5	72.3	59.7	64.2	11.8	116.78M	6.6
LFG-Net*	63.9	90.1	76.8	63.6	69.5	42.5	174.28M	5.0

Performance on AirSARShip

Ship detection and ship instance segmentation results on AirSARShip dataset. The models are trained for 36 epochs with the initial learning rate at 0.0025. In addition to the MS COCO evaluation metrics, Parameters, and FPS, we also provide the gap between AP<sup>Bbox</sup> and AP<sup>Mask</sup>.

Model	AP<sup>Bbox</sup>	AP<sub>50</sub>	AP<sub>75</sub>	AP<sub>S</sub>	AP<sub>M</sub>	AP<sub>L</sub>	AP<sup>Mask</sup>	AP<sub>50</sub>	AP<sub>75</sub>	AP<sub>S</sub>	AP<sub>M</sub>	AP<sub>L</sub>	Gap	Params.	FPS
Mask R-CNN	56.8	82.2	64.0	49.4	61.9	25.7	49.1	77.1	56.9	40.1	53.3	30.8	7.7	43.75M	21.8
Point Rend	58.3	83.4	67.1	50.0	63.7	29.7	54.1	80.5	64.0	41.6	59.0	40.9	4.2	55.53M	20.1
GRoIE	57.5	82.0	66.3	49.2	62.9	28.2	51.4	78.7	59.9	40.7	55.8	37.2	6.1	47.54M	10.3
Mask Scoring R-CNN	58.0	83.1	66.1	55.0	63.0	32.2	49.4	77.6	56.5	39.3	53.6	34.0	8.6	60.01M	20.8
R-ARE-Net	56.6	83.3	64.8	49.0	61.9	31.5	53.8	80.4	63.8	46.4	58.5	32.2	2.8	46.58M	12.1
QueryInst	40.1	64.5	42.8	37.3	43.1	25.4	35.4	60.3	38.3	29.1	38.4	31.7	4.7	172.22M	6.4
Cascade Mask R-CNN	60.6	83.3	69.2	50.8	66.0	34.1	50.9	78.4	58.5	40.0	55.4	34.3	9.7	76.80M	18.0
Hybrid Task Cascade	60.7	84.1	69.0	50.9	66.1	36.3	52.7	80.2	61.3	41.5	57.0	39.8	8.0	79.73M	13.6
Detectors	61.7	85.0	69.2	51.5	67.1	37.5	54.5	81.5	63.6	42.6	58.9	42.6	7.2	134.00M	7.7
SCNet	60.1	83.2	67.7	50.8	65.6	32.9	54.3	80.6	63.5	42.8	58.7	42.5	5.8	94.29M	10.5
LFG-Net*	64.8	84.1	73.6	58.8	69.3	39.2	61.8	82.1	70.8	53.8	65.3	53.4	3.0	174.28M	9.0

Citation

If the project helps your research, please cite our paper:

@article{wei2022lfgnet,
title={LFG-Net: Low-level Feature Guided Network for Precise Ship Instance Segmentation in SAR Images},
author={Wei Shunjun, Zeng Xiangfeng, Zhang Hao, Zhou Zichen, Shi Jun, Zhang Xiaoling},
journal={IEEE Transactions on Geoscience and Remote Sensing},
volume={60},
year={2022},
publisher={IEEE}
}