This project is based on VINS-FUSION. I've merged superpoint、lightglue and MixVPR into VINS-FUSION for loop closure.If it is helpful to you ,please consider starring this repository.

If you use D_VINS for your academic research, please cite our related papers. pdf

• Shi, Y.; Li, R.; Shi, Y.; Liang, S. A Robust and Lightweight Loop Closure Detection Approach for Challenging Environments. Drones 2024, 8, 322. https://doi.org/10.3390/drones8070322

1、Env

python==3.8.18

onnx==1.13.1

onnxruntime==1.16.3

torch==1.13.1+cu117 and torch2.1+cu12.1

cuda == 11.7

TensorRT == 8.6.1.6

faiss == 1.7.2

opencv == 3.4.10

2070 SUPER

2、Build Vins-Fusion

please refer to Vins-Fusion

3、export onnx

• mixvpr

    model = VPRModel(backbone_arch='resnet50',
                     layers_to_crop=[4],
                     agg_arch='MixVPR',
                     agg_config={'in_channels': 1024,
                                 'in_h': 20,
                                 'in_w': 20,
                                 'out_channels': 256,
                                 'mix_depth': 4,
                                 'mlp_ratio': 1,
                                 'out_rows': 2},
                     )

    state_dict = torch.load('/home/sy/sy/Mix_ws/src/mixvpr/model/resnet50_MixVPR_512_channels(256)_rows(2).ckpt')
    model.load_state_dict(state_dict)
    model.eval()
    model.cpu()

    input = torch.randn(1, 3, 320, 320)
    output_path = '/home/sy/sy/Mix_ws/src/mixvpr/mix_512.onnx'

    torch.onnx.export(model, input, output_path,
                      verbose=False,
                      opset_version=13,
                      input_names=['img'],
                      output_names=['des'],
                      )

    model_sim, flag = onnxsim.simplify(output_path)
    if flag:
        onnx.save(model_sim, output_path)
        print("---------simplify onnx successfully---------")
    else:
        print("---------simplify onnx failed-----------")

• SP + LG

1、Download the weights from here superpoint_weights

2、Follow the guidence of lightglue-ONNX to export LG onnx in torch2.1+cuda12.1, and I suggest you to export with my params if you dont want to change any codes. Also, you can directly download the superpoint_512.onnx provided in model v0.1.3

See issue3 and issue 11 for more imformation

--dynamic
--extractor_type=superpoint
--extractor_path=YOUR_SP_ONNX_PATH
--lightglue_path=YOUR_LG_ONNX_PATH
--max_num_keypoints=YOUR_NUM / MY_NUM = 512

example :

if __name__ == "__main__":
    #generate onnx model
    args = parse_args()
    export_onnx(**vars(args))

    #simplify your onnx model
	#replace it with your exsited onnx path
    output_path_sp=('/home/sy/sy/Mix_ws/src/mixvpr/model/sp+sg/superpoint_512_new.onnx')

    print('----------start simplifying sp-----------')
    model_sim, flag = onnxsim.simplify(output_path_sp)
    if flag:
        onnx.save(model_sim, output_path_sp)
        print("---------simplify sp successfully---------")
    else:
        print("---------simplify sp failed-----------")
        
    #simplify your onnx model
	#replace it with your exsited onnx path
    output_path_lg=('/home/sy/sy/Mix_ws/src/mixvpr/model/sp+sg/superpoint_512_new.onnx')

    print('----------start simplifying lg-----------')
    model_sim, flag = onnxsim.simplify(output_path_lg)
    if flag:
        onnx.save(model_sim, output_path_lg)
        print("---------simplify lg successfully---------")
    else:
        print("---------simplify lg failed-----------")

• SP_RECOVER

1、Download the weights from here superpoint_weights

2、see ultrapoint.py and follow the example

if __name__ == "__main__":
    model = ultrapoint.UltraPoint().eval()
    model.cpu()
    # 确定好网络输入和输出路径
    #input img_size
    image = torch.randn(1, 1, 480, 752)
    #input kpts.size
    keypoints = torch.randn(1, 512, 2)
    #replace it with yopur path
    output_path = '/home/sy/sy/Mix_ws/src/mixvpr/model/superpoint_recover_des_480x752.onnx'

    # 载入模型，输入、opset版本以及输入输出的名字
    # options
    torch.onnx.export(model,
                      (image,keypoints),
                      output_path,
                      verbose=False,
                      opset_version=17,
                      input_names=["image_r","keypoints_r"],
                      output_names=["scores_r","des_r"],
                      dynamic_axes={
                          "keypoints_r": {1: "num_keypoints"},
                          "des_r": {1: "num_keypoints"},
                          "scores_r": {0: "num_keypoints"},
                      }
                      )
    # 模型简化 simplify your model
    model_sim, flag = onnxsim.simplify(output_path)
    if flag:
        onnx.save(model_sim, output_path)
        print("---------simplify sp_re successfully---------")
    else:
        print("---------simplify sp_re failed-----------")
        

4、将onnx转成engine文件 / transform onnx to .engine

All the engines are transformed in tensorrt8.6.1.6 + cuda11.7 env

Example:

trtexec --onnx='/home/sy/sy/Mix_ws/src/mixvpr/model/sim_800x400_512/superpoint_512.onnx' --fp16 --minShapes=image:1x1x400x800 --optShapes=image:1x1x400x800 --maxShapes=image:1x1x400x800 --saveEngine=/home/sy/sy/Mix_ws/src/mixvpr/model/sim_800x400_512/superpoint_800x400_512.engine --warmUp=500 --duration=10

//SP
trtexec --onnx='/home/sy/sy/Mix_ws/src/mixvpr/model/sim_752x480_512/superpoint_512.onnx'  --fp16 --minShapes=image:1x1x480x752 --optShapes=image:1x1x480x752 --maxShapes=image:1x1x480x752  --saveEngine=/home/sy/sy/Mix_ws/src/mixvpr/model/sim_752x480_512/superpoint_752x480_512.engine --warmUp=500 --duration=10


//LG
trtexec --onnx='/home/sy/sy/lightglue_ws/src/LightGlue-ONNX/weights/my/sim_752x480_1024/superpoint_lightglue.onnx'  --fp16 --saveEngine='/home/sy/sy/Mix_ws/src/mixvpr/model/sim_752x480_1024/superpoint_lightglue_10_1024.engine' --warmUp=500 --duration=10 --minShapes=kpts0:1x10x2,kpts1:1x10x2,desc0:1x10x256,desc1:1x10x256 --optShapes=kpts0:1x512x2,kpts1:1x512x2,desc0:1x512x256,desc1:1x512x256  --maxShapes=kpts0:1x1024x2,kpts1:1x1024x2,desc0:1x1024x256,desc1:1x1024x256


//SP_RE
trtexec --onnx='/home/sy/sy/Mix_ws/src/mixvpr/model/sp_re_752x480_512/superpoint_recover_des_480x752.onnx'  --fp16 --saveEngine='/home/sy/sy/Mix_ws/src/mixvpr/model/sp_re_752x480_512/superpoint_recover_des_480x752.engine' --warmUp=500 --duration=10 --minShapes=keypoints_r:1x20x2  --optShapes=keypoints_r:1x150x2  --maxShapes=keypoints_r:1x512x2


//mixvpr
trtexec --onnx='/home/sy/sy/Mix_ws/src/mixvpr/model/mix1.onnx'  --fp16 --saveEngine=mix1.engine --warmUp=500 --duration=10 

5、performance in some datasets(EuRoC, KAIST urban, 4Seasons)

Our method improves accuracy by 20% over vins-fusion on euroc and even more on difficult datasets such as KAIST urban and 4seasons. It is able to cope with urban highway scenarios and seasonal change scenarios.

使用 / usage

Follow the steps in VINS-Fusion to run this project and follow the format of euroc_stereo_imu_config.yaml to set your own YAML if needed.

matching examples

image size	752*480(euroc)	800*400(4seasons)	1280*560(KAIST urban)
Global Feature Extraction (ms)	1.5	2	2.3
Keyframe Retrival (ms)	0.8	2.2	1.7
Local Feature Extraction(512 points) (ms)	6.1	5.4	10.3
Local Feature Matching (ms)	4.1	4.1	3.8
Total Time cost (ms)	12.5	13.7	18.1
Total Memory cost (MB)	422	400	600

6、TODO

• Replace some functions in this project with kernal function if possible.

7、 Acknowledgements

This project is based on Vins-Fusion

We use ceres solver for non-linear optimization, a generic camera model and GeographicLib.

I use superpoint 、lightglue、MixVPR for loop closure,and the part of TensorRT infrence is based on Linfer

8、 License

The source code is released under GPLv3 license.