DoubangoTelecom comments

Results 44 comments of


                                            DoubangoTelecom

Sobel could be done multiplication-free

And use abs_avg ((abs(gx) + abs(gy) + 1) >> 1) instead of gradient L1 plus scaleAndClip. abs_avg is single instruction in SSE and NEON

_SC_LEVEL1_DCACHE_LINESIZE and _SC_LEVEL1_DCACHE_SIZE not available on Android

Maybe try reading 'devices/system/cpu/cpu0/cache/index0/coherency_line_size'

Brute force matcher, add support for mt based on trains instead of queries when there are more trains than queries

https://github.com/DoubangoTelecom/compv/blob/c8fbe25320515130d13cad2c736f20f2f7cd4316/core/matchers/compv_matcher_bruteforce.cxx#L121

ARM64 MediaPad2 -> MSER ST faster than MSER MT

Also the case for ARM32 on MediaPad2.

ARM64 MediaPad2 -> MSER ST faster than MSER MT

Not the case for Galaxy Tab6 ARM32

AVX-SSE Transition Penalties

For Intrin code

Add support for PRNG using Mersenne Twister

To be used to compute random subsets for Homography (RANSAC)

On many old SIMD code the orphans can be saved without breaking the loops

- Hamming distances (SSE and AVX)

Android native camera from jni

1) Unzip /cygdrive/c/Program Files (x86)/Android/android-sdk/platforms/android-(NUMBER HERE)/android.jar to /cygdrive/c/Projects/tmp 2) go into /cygdrive/c/Projects/tmp/android_jar 3) export PATH=$PATH:"/cygdrive/c/Program Files (x86)/Java/jdk1.7.0_55/bin/" 4) javap -s android/hardware/Camera.class

add "CompVMathDistanceHamming32_Intrin_POPCNT_SSE42" using mula's algorithm

Just like what we're doing with AVX2: CompVMathDistanceHamming32_Intrin_POPCNT_AVX2