This repository includes the implentation of the methods in Representation Based Regression for Object Distance Estimation.

Software environment:

1. Python with the following version and libraries.
python == 3.7.7
tensorflow == 2.1.0
Numpy == 1.18.5
Matplotlib == 3.2.2
SciPy == 1.5.0
scikit-learn == 0.24.1
argparse == 1.1
OpenCV == 4.2.0
pandas == 1.0.4
tqdm == 4.46.1

2. MATLAB -> MATLAB R2019a.

Content:

• Citation
• Getting started with the KITTI Dataset
• Feature Extraction
• Distance Estimation via Representation-based Classification
  • Sparse Representation-based Classification (SRC)
  • Collaborative Representation-based Classification (CRC)
• Distance Estimation using Representation-based Regression (RbR)
  • Convolutional Support Estimator Network (CSEN)
  • Compressive Learning CSEN (CL-CSEN)
• Distance Estimation using Support Vector Regressor (SVR)
• References

If you use method(s) provided in this repository, please cite the following paper:

@misc{ahishali2021representation,
      title={Representation Based Regression for Object Distance Estimation}, 
      author={Mete Ahishali and Mehmet Yamac and Serkan Kiranyaz and Moncef Gabbouj},
      year={2021},
      eprint={2106.14208},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

The left color images of object dataset and the corresponding training labels can be obtained from 3D Object Detection Evaluation 2017.

unzip data_object_image_2.zip
unzip data_object_label_2.zip

After unzipping, move them under kitti-data/ folder and using generate-csv.py script generate a csv file for the KITTI annotations:

python kitti-data/generate-csv.py \
    --input=kitti-data/training/label_2/ \
    --output=kitti-data/annotations.csv

Features are extracted using feature_extraction.py script. The supported models are DenseNet121, VGG19, and ResNet50 (DenseNet-121 [1], VGG19 [2], ResNet-50 [3]):

python feature_extraction.py --model VGG19 

Next, the features are further processed and ordered using processFeatures.m. In the script, please also set the proper model name param.modelName to either DenseNet121, VGG19, or ResNet50 and param.DicDesign to 2D or 1D corresponding to the dictionary designs used in CSEN and CL-CSEN approaches. This procedure is only needed for the CSEN, CL-CSEN, and SVR approaches. If you are only interested in running SRC and CRC methods, you may proceed to the related sections: SRC and CRC. Note that the script of processFeatures.m produces the predicted distances using the CRC-light model that is discussed in the paper.

We formulate the distance estimation task as a representation-based classification problem by estimating the quantized distance values. For example, for the objects between [0.5, 60.5] meters away from the camera, we estimate a quantized distance level from 60 different distance levels ranging between [1, 60] with 1-meter senstivity.

There are implemented 8 different SRC algorithms for the distance estimation task including ADMM [4], Dalm [5], OMP [5], Homotopy [6], GPSR [7], L1LS [8], ℓ1-magic [9], and Palm [5]. You may run all of them in once as follows:

cd src/
run main.m

Alternatively or preferably (e.g., you may choose a specific SRC method since the required run-time is huge for running all SRC methods in once), the selected ones can be defined in the script:

l1method={'solve_ADMM','solve_dalm','solve_OMP','solve_homotopy','solveGPSR_BCm', 'solve_L1LS','solve_l1magic','solve_PALM'}; %'solve_PALM' is very slow

Similarly, please also set the proper model name param.modelName to either DenseNet121, VGG19, or ResNet50.

Distance estimation using the CRC method [10] can be run as follows:

cd crc\
run main.m

The CRC-light model can be run by setting CRC_light = 1 in the script. Please change the model name param.modelName to DenseNet121, VGG19, or ResNet50 to try different features.

Contrary to previous methos, it is possible to directly estimate the object distance information without the quantization step using CSEN and CL-CSEN approaches. As CSEN and CL-CSEN approaches still utilize the representative dictionary, we introduce the term Representation-based Regression (RbR) for the proposed framework.

The CSEN implementation is run as follows:

python regressor_main.py --method CSEN --feature_type DenseNet121

Note that similarly, the feature type can be set to DenseNet121, VGG19, or ResNet50. If you like, only testing can be performed using the provided weights:

python regressor_main.py --method CSEN --feature_type DenseNet121 --weights True

The CL-CSEN implementation is run as follows:

python regressor_main.py --method CL-CSEN --feature_type DenseNet121

The parameter --feature_type can be set to DenseNet121, VGG19, or ResNet50.

Testing of CL-CSEN with the provided weights:

python regressor_main.py --method CL-CSEN --feature_type DenseNet121 --weights True

The SVR method is implemented as a competing regressor. We use the Nystroem method for the kernel approximation since it is unfeasible to compute exact kernel mapping with the given high-dimensional dataset. Hyperparameter search is applied with grid search and then the performance is computed with the found optimal SVR parameters:

python regressor_main.py --method SVR --feature_type DenseNet121

The parameter --feature_type can be set to DenseNet121, VGG19, or ResNet50.

[1] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, "Densely connected convolutional networks," in Proc. IEEE Conf. Comput. Vision and Pattern Recognit. (CVPR), 2017, pp. 4700–4708.
[2] K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition," arXiv preprint arXiv:1409.1556, 2014.
[3] K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proc. IEEE Conf. Comput. Vision and Pattern Recognit. (CVPR), 2016, pp. 770–778.
[4] S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Eckstein et al., "Distributed optimization and statistical learning via the alternating direction method of multipliers," Found. Trends Mach. Learn., vol. 3, no. 1, 2011.
[5] A. Y. Yang, Z. Zhou, A. G. Balasubramanian, S. S. Sastry, and Y. Ma, "Fast l1-minimization algorithms for robust face recognition," IEEE Trans. Image Process., vol. 22, no. 8, pp. 3234–3246, 2013.
[6] D. M. Malioutov, M. Cetin, and A. S. Willsky, "Homotopy continuation for sparse signal representation," in Proc. IEEE Int. Conf. Acoust., Speech, and Signal Process. (ICASSP), vol. 5, 2005, pp. 733–736.
[7] M. A. Figueiredo, R. D. Nowak, and S. J. Wright, "Gradient projection for sparse reconstruction: Application to compressed sensing and other inverse problems," IEEE J. Sel. Topics Signal Process., vol. 1, no. 4, pp. 586–597, 2007.
[8] K. Koh, S.-J. Kim, and S. Boyd, "An interior-point method for large-scale l1-regularized logistic regression," J. Mach. Learn. Res., vol. 8, pp. 1519–1555, 2007.
[9] E. Candes and J. Romberg, "l1-magic: Recovery of sparse signals via convex programming," Caltech, Tech. Rep., 2005. [Online]. Available: https://statweb.stanford.edu/∼candes/software/l1magic/downloads/l1magic.pdf
[10] L. Zhang, M. Yang, and X. Feng, "Sparse representation or collaborative representation: Which helps face recognition?" in Proc. IEEE Int. Conf. Comput. Vision (ICCV), 2011, pp. 471–478.