This repository contains the implementation of the following paper: Learning Soft Estimator of Keypoint Scale and Orientation with Probabilistic Covariant Loss.
This code is developed with Python 3.8 and PyTorch 1.4, but the packages with the later versions should be also suitable. Typically, conda and pip are the recommended ways to configure the environment:
conda create -n S3Esti python=3.8
conda activate S3Esti
conda install numpy matplotlib scipy
conda install pytorch==1.4.0 cudatoolkit=10.1 -c pytorch
pip install imgaug
pip uninstall opencv-python
pip install opencv-contrib-python==3.4.8.29We provide the evaluation code for HPatches dataset, which references the evaluation processes of HesAffNet and POP. Before perform the evaluation on the entire HPatches, you can first verify the environment by running the script directly:
python evaluation.pyIf the environment is configured correctly, the following information will be printed in the terminal:
AffNet_esti_HardNet ---- test_image
v_boat,HA: ...
s_error: ...
AffNet_HardNet ---- test_image
...
Here HesAffNet_HardNet+S3Esti, HesAffNet_HardNet, POP+S3Esti and POP are evaluated on the v_boat sequence in HPatches. Note that the v_boat sequence has been placed in the test_image folder so that the evaluation.py script can be run directly. Furthermore, the pre-trained models of HesAffNet and POP are also included in this repository to simplify the configuration.
In the above process, the statistics_results folder is created automatically and the main statistics results are written in it. After the evaluations of all methods, four text files, namely AffNet_esti_HardNet.txt, POP_esti.txt, AffNet_HardNet.txt, POP.txt, should appear in this folder.
The environment is verified to be correct if the above statistics results can be outputted. Then you can place other data in the hpatches-sequences-release folder to further evaluate the methods. Note the format of the data should be consistent to the HPatches sequences. One convenient way is to download the HPatches sequences first, and then unzip it into the hpatches-sequences-release folder.
For more details about all parameters of evaluation.py, run python evaluation.py --help.
Before drawing the accuracy curves, the evaluation in the last step should have been finished. Then the accuracy curves can be obtained by running the script:
python draw_result_line_easy_hard.pyThe accuracy curves will appear in the figure_results/ folder. The default curves are drawn for the homography accuracy metric. For more details about all parameters of draw_result_line_easy_hard.py, run python draw_result_line_easy_hard.py --help.
You can first verify the environment by running the script directly:
python train.pyThis command performs the training process, and the training images are in the demo_input_images/ folder. If the environment is configured correctly, the following information will be printed in the terminal:
ep:3, iter:0/1, l:x.xxx, l_s:x.xxx, l_a:x.xxx, s_e:x.xxxx, a_e:x.xxxx, ...
ep:7, iter:0/1, l:x.xxx, l_s:x.xxx, l_a:x.xxx, s_e:x.xxxx, a_e:x.xxxx, ...
...
And the training process will write the checkpoints into the S3Esti_checkpoint/ folder. The name of the checkpoint is formatted as checkpoint_end_ep_x.
The environment is verified to be correct if the above process can be finished without error. To train your model, you can place other data in the demo_input_images folder, or specify the training folder with the --train-image-path parameter:
python train.py --training-path /the/path/of/training/datasetTo reproduce the performance in the paper, you can train the model with COCO 2014 training set (containing 82783 images). In our experiments, about 30 epochs are generally required to achieve the performance similar to that in the paper.
You can also set the --restore-checkpoint-path parameter to make the model be initialized with the given checkpoint:
python train.py --restore-checkpoint-path /the/path/of/checkpointFor more details about all parameters of train.py, run python train.py --help.