speed_and_accuracy.txt

SiftGPU is an implementation of SIFT [1] for GPU. SiftGPU processes pixels parallely to build Gaussi

SPEED

1. The first several rounds of SiftGPU might be slow, don't worry. If it is always slow, the reason
 might be either the graphic card is old or the graphic memory is small. 

2. Texture reallocation happens when a new image does not fit in the current allocated storage. 
To get less reallocations, you can pre-allocate storage to fit the largest image size, or just 
use images with the same size. 

3. Loading some compressed images (.e.g jpg) may take a lot of time on decompressing. Using binary 
pgm files or directly specifying data in memory can achive better performance. Writing ASCII 
SIFT files is also slow. 

4. The packd version saves GPU memory but also run faster than the unpacked, which is default now. 

5. SiftGPU might be faster with older grpahic card drivers than with newer ones. 

6. The descriptor normalization in the OpenGL-based implementation is running on CPU. New versions 
are now using SSE, which improves the speed for this part a lot. 

7. The orientation computation in unpacked implementation is occasionally slow under single 
orientation  mode (-m 1) or packed orientation mode (-m2p). By default, siftgpu uses 2 orientations
 (-m 2), which should be fine. It is still unresolved.


8. The thread block settings in the CUDA-based SiftGPU are currently tuned for my GPU nVidia GTX 
8800, which may not be optimized for other GPUs.


ACCURACY


1. The latest version of SiftGPU now has comparable accuracy with CPU implementatins. Evaluation on 
box.pgm of Lowe's package now gives around 600 matches, which is close to SIFT++.

2. In orientation computation, SiftGPU uses a factor 2.0 * sigma as the sample window size, which is 
smaller than the typical value 3.0. Changing it from 2.0 to 3.0 reduces the speed of this step by %40, but 
gives only a very small improvements in matching. You can change it by specifying parameter "-w 3". 

3. In keypoint localization, SiftGPU refines the location only once by default. You can change it by 
specifiying "-s n", to get n iterations (Only available in cg unpacked implementation). SiftGPU does 
not move the level of keypoints in the refinement. 

4. The feature locations are having a (0.5,0.5) offset compared with CPU implementations by default.  
(0, 0) in texture is at the top-left coorner (instead of center) of the top-left pixel. You can use 
the center as (0, 0) by specifying "-loweo"

5. By default, SiftGPU does not do Upsampling(-fo -1), To match it with Lowe's implementation you need
 to use "-fo -1 -loweo". 

6. SiftGPU may get slightly different result on different GPUs due to different floating point precision. 
SiftGPU is tested on limited types of graphic cards/OS, working on your graphic card is not guaranteed. 

IF it returns different number of features at different run on the same card, then something is going wrong, 
and probably some special tricks need to be used. Please email me if that happens. 

7. Please look at the saved SIFT file to make sure there are no weired descriptors( for example, all
of the numbers of a descriptor are 45, or any number in a descriptor is larger than 255)


KWOWN ISSUES.
1. SiftGPU may have problem with dual monitor.
2. Slow on 7950. Changing GlobalParam::_iTexFormat to GL_RGBA16F_ARB can make it work. Unknown reason.
3. Experiments on 8600 show problems. It works fine for the first image, but gets wrong keypoints after.