用自己的数据训练LoFTR，这可能是全网你唯一能找到的教程了

This repo is creation for training LoFTR on custom dataset

This is the initial LoFTR README file.

As we all know, training LoFTR is not easy, I mean the official dont offer too much info for training our own custom datasets, so I give it a try.

This repo finishs the following things:

• Using Colmap for sparse reconstruction.

• Using Colmap for dense reconstruction and generate depth map.

• Convert depth.bin to depth.h5 file.

• Convert camera params to .npz file.

• Automatically Generate train data structure as MegaDepth format.

• Modify a little code because of the errors in the initial LoFTR.

Github Repo: training LoFTR on custom dataset

Environment

you should follow the README file of LoFTR to prepare the Environment.

conda env create -f environment.yaml
conda activate loftr
pip install -r requirements_loftr.txt

requirements_loftr.txt is my environment.

Prepare your own data

1.Using Colmap for sparse reconstruction and dense reconstruction

First of all, you should install Colmap (best cuda version).

If you have your own COLMAP dataset without undistortion (e.g., using OPENCV camera), you can try to just run the last part of the script:
Put the images in input and the COLMAP info in a subdirectory distorted:

<location>
|---input
|   |---<image 0>
|   |---<image 1>
|   |---...

Then run

python generate_train_data/colmap_convert.py -s <location> [--skip_matching] [--resize] #If not resizing, ImageMagick is not needed
 
# the input dir is in the <location>

After running the command, you will get the sparse reconstruction and the depth maps.

<location>
|---input
|   |---<image 0>
|   |---<image 1>
|   |---...
|---images
|   |---<image 0>
|   |---<image 1>
|   |---...
|---distorted
    |---database.db
    |---sparse
        |---0
            |---...
|---sparse
|   |---0
        |---cameras.bin
        |---images.bin
        |---points3D.bin
        |---cameras.txt
        |---images.txt
        |---points3D.txt
    |---cameras.bin
    |---images.bin
    |---points3D.bin
|---stereo
    |---depth_maps
        |---image_0.photometric.bin
        |---image_1.photometric.bin
        |---...
    |---normal_maps
        |---image_0.photometric.bin
        |---image_1.photometric.bin
        |---...

2.Converting depth.bin to depth.h5 file

Then run the following command to convert the depth.bin to depth.h5 file:

python generate_train_data/read_depth.py \
    --depth_map_dir YOUR_COLMAP_DEPTH_MAP_DIR \
    --normal_map_dir YOUR_COLMAP_NORMAL_MAP_DIR \
    --output_dir YOUR_OUTPUT_h5_DEPTH_DIR

Then you will get the following data dir:

<location>
|---input
|   |---<image 0>
|   |---<image 1>
|   |---...
|---images
|   |---<image 0>
|   |---<image 1>
|   |---...
|---depths
|   |---<image 0>.h5
|   |---<image 1>.h5
|   |---...
|---depth_image
|   |---<image 0>
|   |---<image 1>
|   |---...
|---distorted
    |---database.db
    |---sparse
        |---0
            |---...
|---sparse
|   |---0
        |---cameras.bin
        |---images.bin
        |---points3D.bin
        |---cameras.txt
        |---images.txt
        |---points3D.txt
    |---cameras.bin
    |---images.bin
    |---points3D.bin
|---stereo
    |---depth_maps
        |---image_0.photometric.bin
        |---image_1.photometric.bin
        |---...
    |---normal_maps
        |---image_0.photometric.bin
        |---image_1.photometric.bin
        |---...

depths dir contain the following files:

depth_image dir contain the following files:

3.Automatically generate the train data as MegaDepth format

Then run the following command to generate train data:

python generate_train_data/preprocess_scene.py --base_path YOUR_BASE_PATH --scene_name YOUR_SCENE_NAME

Then you will get the following data dir:

<location>
|---input
|   |---<image 0>
|   |---<image 1>
|   |---...
|---images
|   |---<image 0>
|   |---<image 1>
|   |---...
|---depths
|   |---<image 0>.h5
|   |---<image 1>.h5
|   |---...
|---depth_image
|   |---<image 0>
|   |---<image 1>
|   |---...
|---sparse
|   |---0
        |---cameras.bin
        |---images.bin
        |---points3D.bin
        |---cameras.txt
        |---images.txt
        |---points3D.txt
    |---cameras.bin
    |---images.bin
    |---points3D.bin
|---<scene_name>
|   |---index
|       |---scene_info
|           |---<scenen_name>_train.npz
|           |---<scenen_name>_val.npz
|           |---<scenen_name>_test.npz
|       |---trainvaltest_list
|           |---train_list.txt
|           |---val_list.txt
|           |---test_list.txt
|   |---test
|       |---depths
|           |---image0.h5
|           |---image1.h5
|           |---...
|       |---images
|           |---image0.png
|           |---image1.png
|           |---...
|   |---train
|       |---depths
|           |---image0.h5
|           |---image1.h5
|           |---...
|       |---images
|           |---image0.png
|           |---image1.png
|           |---...
|   |---val
|       |---depths
|           |---image0.h5
|           |---image1.h5
|           |---...
|       |---images
|           |---image0.png
|           |---image1.png
|           |---...

If you want to combine the val and test data, you should move the val/depths/* and val/images/* to test/depths* and test/images respectively.

And just modify the index/trainvaltest_list/test_list.txt, as the following shows:

test_list.txt

<scene_name>_test
<scene_name>_val

Finally move the <scene_name> dir to LoFTR_Project/data

4.Modify Config file

Using the template creating your data config file. And copy the config to the LoFTR_Project/configs/data dir.

change YOUR_DATA_SCENE

from configs.data.base import cfg

TRAIN_BASE_PATH = "data/YOUR_DATA_SCENE/index"
cfg.DATASET.TRAINVAL_DATA_SOURCE = "MegaDepth"
cfg.DATASET.TRAIN_DATA_ROOT = "data/YOUR_DATA_SCENE/train"
cfg.DATASET.TRAIN_NPZ_ROOT = f"{TRAIN_BASE_PATH}/scene_info"
cfg.DATASET.TRAIN_LIST_PATH = f"{TRAIN_BASE_PATH}/trainvaltest_list/train_list.txt"
cfg.DATASET.MIN_OVERLAP_SCORE_TRAIN = 0.0

TEST_BASE_PATH = "data/YOUR_DATA_SCENE/index"
cfg.DATASET.TEST_DATA_SOURCE = "MegaDepth"
cfg.DATASET.VAL_DATA_ROOT = cfg.DATASET.TEST_DATA_ROOT = "data/YOUR_DATA_SCENE/test"
cfg.DATASET.VAL_NPZ_ROOT = cfg.DATASET.TEST_NPZ_ROOT = f"{TEST_BASE_PATH}/scene_info"
cfg.DATASET.VAL_LIST_PATH = cfg.DATASET.TEST_LIST_PATH = f"{TEST_BASE_PATH}/trainvaltest_list/test_list.txt"
cfg.DATASET.MIN_OVERLAP_SCORE_TEST = 0.0   # for both test and val

# 368 scenes in total for MegaDepth
# (with difficulty balanced (further split each scene to 3 sub-scenes))
cfg.TRAINER.N_SAMPLES_PER_SUBSET = 100

cfg.DATASET.MGDPT_IMG_RESIZE = 640  # for training on 11GB mem GPUs

5.Full command

# 1. first step
python generate_train_data/colmap_convert.py -s <location> 

# 2. second step
python generate_train_data/read_depth.py \
    --depth_map_dir <location>/stereo/depth_maps \
    --normal_map_dir <location>/stereo/normal_maps \
    --output_dir <location>

# 3. third step
python generate_train_data/preprocess_scene.py --base_path <location> --scene_name <scene_name>

# change <location> and <scene_name> for your data

Then you could train your own LoFTR model happily!😊