Jaco Cronje - Academia.edu (original) (raw)
Papers by Jaco Cronje
We propose a fast local image feature detector and descriptor that is implementable on the GPU. O... more We propose a fast local image feature detector and descriptor that is implementable on the GPU. Our method is the first GPU implementation of the popular FAST detector. A simple but novel method of feature orientation estimation which can be calculated in constant time is proposed. The robustness and reliability of our orientation estimation is validated against rotation invariant descriptors such as SIFT and SURF. Furthermore, we propose a binary feature descriptor which is robust to noise, scalable, rotation invariant, fast to compute in parallel and maintains low memory consumption. The proposed method demonstrates good robustness and very fast computation times, making it usable in real-time applications.
Proceedings of SPIE, May 13, 2011
Inverse lens distortion modelling allows one to find the pixel in a distorted image which corresp... more Inverse lens distortion modelling allows one to find the pixel in a distorted image which corresponds to a known point in object space, such as may be produced by RADAR. This paper extends recent work using neural networks as a compromise between processing complexity, memory usage and accuracy. The already encouraging results are further enhanced by considering different neuron activation functions, architectures, scaling methodologies and training techniques.
Proceedings of SPIE, May 20, 2013
This work extends earlier work on the real-time photogrammetric stitching of staring arrays of hi... more This work extends earlier work on the real-time photogrammetric stitching of staring arrays of high resolution videos on commercial off the shelf hardware. The blending is both further optimised for Graphics Processor Unit (GPU) implementation and extended from one to two dimensions to allow for multiple layers or arbitrary arrangements of cameras. The incorporation of stabilisation inputs allows the stitching algorithm to provide space stabilised panoramas. The final contribution is to decrease the sensitivity to depth of the stitching procedure, especially for wide aperture baselines. Finally timing tests and some resultant stitched panoramas are presented and discussed.
This paper presents a comparison of several established and recent image feature-descriptors to r... more This paper presents a comparison of several established and recent image feature-descriptors to register long wave infra-red images in the 8-14 µm band to visual band images. The feature descriptors were chosen to include robust algorithms, SURF and SIFT-and fast algorithms, BRISK and BFROST. To evaluate the feature-descriptors a ground truth was created by determining the intrinsic and extrinsic camera calibration parameters for the cameras and using this to photogrammetrically relate pixel positions between the images. The inlier results of each feature descriptor for the top 20%, 50% and 100% of the matches (based on match strength) were used to create a homography. The average pixel error between the homography reprojected feature points and the photogrammetric reprojection was used as the error. The results show that none of the descriptors perform well in standard form, with BFROST faring slightly better than the other algorithms. This suggests a need to modify the algorithms to detect physical/structural features and de-emphasise textural features.
The work in this paper proposes an approach to reduce over-fitting when training a convolutional ... more The work in this paper proposes an approach to reduce over-fitting when training a convolutional neural network (CNN) by applying class based data augmentation. The classification task of automatically detecting drivers engaged in distracted behaviors seen from a dashboard camera is addressed. Firstly, important image key-points are detected and their regions extracted. Thereafter, the extracted regions are combined into a single image and class based data augmentation is applied to the different extracted regions. Class based data augmentation randomly interchange regions between different images of the same class. The results show that this approach improves the generalization of the network and increases the variability of the data set which reduces over-fitting.
Proceedings of the Annual Conference of the South African Institute of Computer Scientists and Information Technologists on - SAICSIT '16, 2016
The US Patent and Trademark Office, together with the NASA Tournament Lab, launched a contest to ... more The US Patent and Trademark Office, together with the NASA Tournament Lab, launched a contest to develop specialized algorithms to help bring the seven million patents presently in the patent archive into the digital age. The contest was hosted by TopCoder.com, the largest competitive online software developer community. The challenge was to detect, segment and recognize figures, captions and part labels from patent drawing images. The solution presented in this work was the winning submission.
SPIE Proceedings, 2011
Inverse lens distortion modelling allows one to find the pixel in a distorted image which corresp... more Inverse lens distortion modelling allows one to find the pixel in a distorted image which corresponds to a known point in object space, such as may be produced by a RADAR. This paper extends recent work using neural networks as a compromise between processing complexity, memory usage and accuracy. The already encouraging results are further enhanced by considering different neuron activation functions, architectures, scaling methodologies and training techniques. The errors are given in terms of microns on the detector to facilitate fair comparison between different resolutions and pixel sizes.
Automatic Target Recognition XXIII, 2013
Abstract—This paper presents a comparison of several es-tablished and recent image feature-descri... more Abstract—This paper presents a comparison of several es-tablished and recent image feature-descriptors to register long wave infra-red images in the 8–14 µm band to visual band images. The feature descriptors were chosen to include robust algorithms, SURF and SIFT — and fast algorithms, BRISK and BFROST. To evaluate the feature-descriptors a ground truth was created by determining the intrinsic and extrinsic camera calibration parameters for the cameras and using this to photogrammetrically relate pixel positions between the images. The inlier results of each feature descriptor for the top 20%, 50 % and 100 % of the matches (based on match strength) were used to create a homography. The average pixel error between the homography reprojected feature points and the photogrammetric reprojection was used as the error. The results show that none of the descriptors perform well in standard form, with BFROST faring slightly better than the other algorithms. This suggests a need to modify t...
23rd Annual Symposium of the Pattern Recognition Association of South Africa (PRASA), Pretoria, S... more 23rd Annual Symposium of the Pattern Recognition Association of South Africa (PRASA), Pretoria, South Africa, 29-30 November 2012. Published in PRASA 2012.
Proceedings of the 9th International Conference on Computer and Automation Engineering - ICCAE '17
The work in this paper proposes an approach to reduce over-fitting when training a convolutional ... more The work in this paper proposes an approach to reduce over-fitting when training a convolutional neural network (CNN) by applying class based data augmentation. The classification task of automatically detecting drivers engaged in distracted behaviors seen from a dashboard camera is addressed. Firstly, important image key-points are detected and their regions extracted. Thereafter, the extracted regions are combined into a single image and class based data augmentation is applied to the different extracted regions. Class based data augmentation randomly interchange regions between different images of the same class. The results show that this approach improves the generalization of the network and increases the variability of the data set which reduces over-fitting.
2015 International Conference on Image and Vision Computing New Zealand (IVCNZ)
Automatic Target Recognition XXIII, 2013
ABSTRACT This work extends earlier work on the real-time photogrammetric stitching of staring arr... more ABSTRACT This work extends earlier work on the real-time photogrammetric stitching of staring arrays of high resolution videos on commercial off the shelf hardware. The blending is both further optimised for Graphics Processor Unit (GPU) implementation and extended from one to two dimensions to allow for multiple layers or arbitrary arrangements of cameras. The incorporation of stabilisation inputs allows the stitching algorithm to provide space stabilised panoramas. The final contribution is to decrease the sensitivity to depth of the stitching procedure, especially for wide aperture baselines. Finally timing tests and some resultant stitched panoramas are presented and discussed.
Visual Information Processing XX, 2011
Inverse lens distortion modelling allows one to find the pixel in a distorted image which corresp... more Inverse lens distortion modelling allows one to find the pixel in a distorted image which corresponds to a known point in object space, such as may be produced by a RADAR. This paper extends recent work using neural networks as a compromise between processing complexity, memory usage and accuracy. The already encouraging results are further enhanced by considering different neuron activation functions, architectures, scaling methodologies and training techniques. The errors are given in terms of microns on the detector to facilitate fair comparison between different resolutions and pixel sizes. .za * openMP (www.openMP.org) is API for multi-platform shared-memory parallel programming in C/C++ and Fortran.
We propose a fast local image feature detector and descriptor that is implementable on the GPU. O... more We propose a fast local image feature detector and descriptor that is implementable on the GPU. Our method is the first GPU implementation of the popular FAST detector. A simple but novel method of feature orientation estimation which can be calculated in constant time is proposed. The robustness and reliability of our orientation estimation is validated against rotation invariant descriptors such as SIFT and SURF. Furthermore, we propose a binary feature descriptor which is robust to noise, scalable, rotation invariant, fast to compute in parallel and maintains low memory consumption. The proposed method demonstrates good robustness and very fast computation times, making it usable in real-time applications.
Proceedings of SPIE, May 13, 2011
Inverse lens distortion modelling allows one to find the pixel in a distorted image which corresp... more Inverse lens distortion modelling allows one to find the pixel in a distorted image which corresponds to a known point in object space, such as may be produced by RADAR. This paper extends recent work using neural networks as a compromise between processing complexity, memory usage and accuracy. The already encouraging results are further enhanced by considering different neuron activation functions, architectures, scaling methodologies and training techniques.
Proceedings of SPIE, May 20, 2013
This work extends earlier work on the real-time photogrammetric stitching of staring arrays of hi... more This work extends earlier work on the real-time photogrammetric stitching of staring arrays of high resolution videos on commercial off the shelf hardware. The blending is both further optimised for Graphics Processor Unit (GPU) implementation and extended from one to two dimensions to allow for multiple layers or arbitrary arrangements of cameras. The incorporation of stabilisation inputs allows the stitching algorithm to provide space stabilised panoramas. The final contribution is to decrease the sensitivity to depth of the stitching procedure, especially for wide aperture baselines. Finally timing tests and some resultant stitched panoramas are presented and discussed.
This paper presents a comparison of several established and recent image feature-descriptors to r... more This paper presents a comparison of several established and recent image feature-descriptors to register long wave infra-red images in the 8-14 µm band to visual band images. The feature descriptors were chosen to include robust algorithms, SURF and SIFT-and fast algorithms, BRISK and BFROST. To evaluate the feature-descriptors a ground truth was created by determining the intrinsic and extrinsic camera calibration parameters for the cameras and using this to photogrammetrically relate pixel positions between the images. The inlier results of each feature descriptor for the top 20%, 50% and 100% of the matches (based on match strength) were used to create a homography. The average pixel error between the homography reprojected feature points and the photogrammetric reprojection was used as the error. The results show that none of the descriptors perform well in standard form, with BFROST faring slightly better than the other algorithms. This suggests a need to modify the algorithms to detect physical/structural features and de-emphasise textural features.
The work in this paper proposes an approach to reduce over-fitting when training a convolutional ... more The work in this paper proposes an approach to reduce over-fitting when training a convolutional neural network (CNN) by applying class based data augmentation. The classification task of automatically detecting drivers engaged in distracted behaviors seen from a dashboard camera is addressed. Firstly, important image key-points are detected and their regions extracted. Thereafter, the extracted regions are combined into a single image and class based data augmentation is applied to the different extracted regions. Class based data augmentation randomly interchange regions between different images of the same class. The results show that this approach improves the generalization of the network and increases the variability of the data set which reduces over-fitting.
Proceedings of the Annual Conference of the South African Institute of Computer Scientists and Information Technologists on - SAICSIT '16, 2016
The US Patent and Trademark Office, together with the NASA Tournament Lab, launched a contest to ... more The US Patent and Trademark Office, together with the NASA Tournament Lab, launched a contest to develop specialized algorithms to help bring the seven million patents presently in the patent archive into the digital age. The contest was hosted by TopCoder.com, the largest competitive online software developer community. The challenge was to detect, segment and recognize figures, captions and part labels from patent drawing images. The solution presented in this work was the winning submission.
SPIE Proceedings, 2011
Inverse lens distortion modelling allows one to find the pixel in a distorted image which corresp... more Inverse lens distortion modelling allows one to find the pixel in a distorted image which corresponds to a known point in object space, such as may be produced by a RADAR. This paper extends recent work using neural networks as a compromise between processing complexity, memory usage and accuracy. The already encouraging results are further enhanced by considering different neuron activation functions, architectures, scaling methodologies and training techniques. The errors are given in terms of microns on the detector to facilitate fair comparison between different resolutions and pixel sizes.
Automatic Target Recognition XXIII, 2013
Abstract—This paper presents a comparison of several es-tablished and recent image feature-descri... more Abstract—This paper presents a comparison of several es-tablished and recent image feature-descriptors to register long wave infra-red images in the 8–14 µm band to visual band images. The feature descriptors were chosen to include robust algorithms, SURF and SIFT — and fast algorithms, BRISK and BFROST. To evaluate the feature-descriptors a ground truth was created by determining the intrinsic and extrinsic camera calibration parameters for the cameras and using this to photogrammetrically relate pixel positions between the images. The inlier results of each feature descriptor for the top 20%, 50 % and 100 % of the matches (based on match strength) were used to create a homography. The average pixel error between the homography reprojected feature points and the photogrammetric reprojection was used as the error. The results show that none of the descriptors perform well in standard form, with BFROST faring slightly better than the other algorithms. This suggests a need to modify t...
23rd Annual Symposium of the Pattern Recognition Association of South Africa (PRASA), Pretoria, S... more 23rd Annual Symposium of the Pattern Recognition Association of South Africa (PRASA), Pretoria, South Africa, 29-30 November 2012. Published in PRASA 2012.
Proceedings of the 9th International Conference on Computer and Automation Engineering - ICCAE '17
The work in this paper proposes an approach to reduce over-fitting when training a convolutional ... more The work in this paper proposes an approach to reduce over-fitting when training a convolutional neural network (CNN) by applying class based data augmentation. The classification task of automatically detecting drivers engaged in distracted behaviors seen from a dashboard camera is addressed. Firstly, important image key-points are detected and their regions extracted. Thereafter, the extracted regions are combined into a single image and class based data augmentation is applied to the different extracted regions. Class based data augmentation randomly interchange regions between different images of the same class. The results show that this approach improves the generalization of the network and increases the variability of the data set which reduces over-fitting.
2015 International Conference on Image and Vision Computing New Zealand (IVCNZ)
Automatic Target Recognition XXIII, 2013
ABSTRACT This work extends earlier work on the real-time photogrammetric stitching of staring arr... more ABSTRACT This work extends earlier work on the real-time photogrammetric stitching of staring arrays of high resolution videos on commercial off the shelf hardware. The blending is both further optimised for Graphics Processor Unit (GPU) implementation and extended from one to two dimensions to allow for multiple layers or arbitrary arrangements of cameras. The incorporation of stabilisation inputs allows the stitching algorithm to provide space stabilised panoramas. The final contribution is to decrease the sensitivity to depth of the stitching procedure, especially for wide aperture baselines. Finally timing tests and some resultant stitched panoramas are presented and discussed.
Visual Information Processing XX, 2011
Inverse lens distortion modelling allows one to find the pixel in a distorted image which corresp... more Inverse lens distortion modelling allows one to find the pixel in a distorted image which corresponds to a known point in object space, such as may be produced by a RADAR. This paper extends recent work using neural networks as a compromise between processing complexity, memory usage and accuracy. The already encouraging results are further enhanced by considering different neuron activation functions, architectures, scaling methodologies and training techniques. The errors are given in terms of microns on the detector to facilitate fair comparison between different resolutions and pixel sizes. .za * openMP (www.openMP.org) is API for multi-platform shared-memory parallel programming in C/C++ and Fortran.