doubangotelecom / compv Goto Github PK

Insanely fast Open Source Computer Vision library for ARM and x86 devices (Up to #50 times faster than OpenCV)

License: GNU General Public License v3.0

Assembly 8.56% C++ 50.35% C 40.24% Objective-C 0.13% Shell 0.01% CMake 0.39% Java 0.11% Perl 0.18% Objective-C++ 0.02% Batchfile 0.01% SWIG 0.01%

computer-vision artificial-intelligence artificial-neural-networks machine-learning ocr-recognition text-recognition neural-network support-vector-machines boosting adaboost

compv's Introduction

This is our R&D Computer Vision project. For now there is no documentation and it's not intended to be used by other companies.

Speed comparison against OpenCV: speed_compare (core i7 quad@4ghz).txt

Insanely faaaaast

Hand-written assembler (10% of the code)
SIMD (SSE, AVX, NEON)
GPGPU (CUDA, OpenVINO, OpenCL, OpenGL, NNAPI and Metal)
Smart multithreading (minimal context switch, no false-sharing, no boundaries crossing..)
Smart memory access (data alignment, cache pre-load, cache blocking, non-temporal load/store for minimal cache pollution, smart reference counting...)
Fixed-point math
... and many more

SDKs using our code

Online demo apps using our code

Technical questions

Please check our discussion group or twitter account

compv's People

Stargazers

Watchers

compv's Issues

cache optimisations and many other stuffs (e.g cupid) must be tested on AMD CPUs

COMPV_INTERLEAVE_I8_XMM_SSE2, reverse line 2/3

Use letsencrypt.org to secure compv.org

CompVThread::setAffinity not implemented on OSX and iOS

Most of the asm code could be opti for x64 (remove, pushs, remove allocation, remove saving/restoring xmm/ymm registers...)

reallocAligned doesn't copy old memory to the newly allocated one

Add FastStrenghts32 implementation using AVX-only code and VPHMINPOSUW

COMPV_YASM_SAVE_RSP_TO_MM0, COMPV_YASM_RESTORE_RSP_FROM_MM0

32bit -> movd
64bit -> movq

FAST9 and FAST12 using ragel state machine

False sharing issue on interest points vector for FAST

RGBA->I420 with odd width doesn't work or YUV player is buggy

Cache alignment issue on all SIMD code

xmm7 loading strides

Make box sorting multi-threaded

use vzeroall instead of vzeroupper and omit vpxor ymm, ymm

Always use 'lea edi, [edi +4]' instead of 'add edi, 4'

CompVThread::setAffinity not tested on Android

Replace COMPV_YASM_SAVE_XMM with COMPV_YASM_SAVE_YMM in AVX code

implement fastArctan2 for embedded devices

http://www.ntu.edu.sg/home/aukil/papers/conf/2011_IEEE-ISIE11_Fast-arctan.pdf

RGBA->I420 C++ version doesn't work

replace all cmovs with setcc when possible. The later is supported on all i386

replace _mm256_permute4x64_epi64 with _mm256_shuffle_epi8

_mm256_permute4x64_epi64 latency = 3
_mm256_shuffle_epi8 latency = 1
Source: https://software.intel.com/sites/landingpage/IntrinsicsGuide

FAST detector: add support for both SAD and maxThreshold in NMS

Add support for PROSAC

http://cmp.felk.cvut.cz/~matas/papers/chum-prosac-cvpr05.pdf

Find efficient way to malloc, realloc, and calloc aligned memory on all platforms

Windows already has these functions

On windows, save XMM6-n registers for avx

Do not use WaitForSingleObject for multithreading: performance issues

Add unit tests with resolution change

compv_popcnt16 is sloow because of the test.

compv_popcnt16_soft is faster and compv_popcnt16_hard is the best

Add support for Apple's accelerate framework

Use "__stdcall" on all ASM extern functions to avoid surprises

Review all x86/x64 assemply using Intel Optimization manual

http://www.intel.com/content/www/us/en/architecture-and-technology/64-ia-32-architectures-optimization-manual.html

Adds support for FMA3 instruction set

Create Doubango feature detector+descriptor

Use Hu's invariant moments (https://en.wikipedia.org/wiki/Image_moment#Rotation_invariant_moments)
Gradient/Histo orient
...

Align the data and stride on the cache line size if it's <128 and pow2(must be but who knows).

This avoid two threads from accessing the same data from the same cache. If the data and stride are aligned on the cache line then, we're sure that we'll not have a lane crossing

Compute function pointers onces and update when CPU flags is update

Only if it doesn't depend on data alignment, e.g. FastDataRow

add support video feature detection and description: motion estimation and compensation

CompVImage::wrap is very slooow

*** Products ***
Object detection, recognition and tracking
Camera calibration tool
Traffic sign recognition
Realtime video stitching
Markerless Augmented-Reality
Google-like StreetView
Text extraction (CCL), recognition (OCR) and loudly reading (Text to speech)
FastANPR: Automatic number-plate recognition
Immersive / Panorama 360 degree video (browser and desktop) - Pan360.org
InDoor location without GPS (Visual Odometry)
Stereo video
Optical Flow
3D reconstruction (SFM - Structure From Motion)
TLD (Track Learn Detect)
Eye tracking
Masques (filtres)
Inverse perspective mapping
WebRTC 360 (JavaScript, C++)
Something like OpenPose

*** BestOf ***
ADAS (traffic sign recognition, traffic light recognition, Brake light detection, straight line detection, Crosswalk detection, speed-detection, LDWS (Lane Departure Warning System), LKAS (Lane Keeping Assistance System), LCW (Lane Changing Warning), DNPW (Do-Not-Pass Warning), road turn prediction, Autonomous cruise control, straight or dashed lines distinction, Forward Collision Warning,
Pedestrian Collision Warning, 360 visio, speed, stoping distance, Eyes off road, Rear camera parking, 360-degree surround view camera for parking, Flashing and stop lights...), Day/Night detector -> https://en.wikipedia.org/wiki/Advanced_driver_assistance_systems