Product wise computer vision development
1 like1,679 views
The document discusses common issues with computer vision (CV) development processes and proposes an improved product-wise approach. Specifically, it notes that CV projects often begin with prototype Matlab code that is not optimized for the target platform and requires rewriting for deployment. This can lead to wasted effort. The document recommends an approach where the initial proof-of-concept uses OpenCV, optimization is considered from the start to minimize memory usage and platform-specific code, and optimized platform-specific OpenCV libraries are later used for each product to simplify porting. Future standardization through APIs like OpenVX is also mentioned.
Product wise computer vision development
- 1. Product-Wise Computer Vision Development
- 2. Agenda • Common CV Development Process • Pitfalls • Why this happens • Suggested Computer Vision Process • Future CV Development • About me
- 3. Common CV Development Initial Develop algorithm on Matlab Demo PC Demo Realtime Demo PC Product PC Product Port Algorithm Port Algorithm Mobile to Platform1 to Platform2 Product
- 4. What’s wrong with it • Matlab code is written as Develop algorithm “demo”/PoC on Matlab • Fast, Dirty • written with low development resource • use proprietary algorithms
- 5. What’s wrong with it Port Algorithm • This step sometime includes: to Platform1 • understanding mobile requirements • Optimization to platform’s SIMD • Memory optimization • sometime full re-written of the algorithm since its too computationally intensive for mobile
- 6. Contributing factors to common CV Project Process • Matlab stage • As a university or a garage SU, developers have low development resources • PC Stage • NIH - Developers prefer their proprietary version over popular algorithms. • at this stage (PoC/Demo) developers doesn't think on mobile optimization
- 7. Contributing factors to common CV Project Process • Platform Stage • Optimization starts from high memory and CP resource demanding PC code • A considerable portion of the code doesn’t use standard OpenCV functions and requires hand made SIMD optimization • Sometimes companies discover that the algorithm is too resource demanding and re-writes the algorithm for the mobile platform
- 8. Product-wise Computer Vision Development Initial Develop initial PoC with OpenCV Demo PC Product PC RT Demo Product Platform fixed-point ported non-OpenCV agnostic Algorithms Minimize Memory requirements Optimized code Use platform-X optimized OpenCV for Platform product (FastCV, TADP, CEVA-CV) Optimized
- 9. Product-wise Computer Vision development • Product wise CV development is guided by the vision of a multi-platform product. This leads to: • Relay on a common CV library->OpenCV • minimization of non-OpenCV functions (saves platform specific SIMD optimization) • Separate between PC Product and optimized C code used as base for porting to mobile / embedded platforms
- 10. Future Development with OpenVX • Khronos compliant SoC will conform to a unified “OpenCV-Like” API • Enables easy porting to all supporting platforms •
- 11. Yossi Cohen - About Me • Android & Video Lecturer • Video Architect & Developer • Computer Vision Consultant • Android Native Developer Yossi Cohen +972545313092 yossicohen19@gmail.com