disparity

for computer vision: disparity

for computer graphic: light map

kMean-2

Iteration = 2

Iteration = 3

Iteration = 4

Iteration = 5

Iteration = 6

Iteration = 7

Iteration = 8

All with k = 6.

kMean

k = 16

k = 8

k = 4

k = 2

All in 2 iterations.

Blur (Image) - Image

Nothing special, I just wanna update my blog.

HDR - 8

FINALLY !!!

HDR - 7

Here is better result by using 3 different response curve (red, green & blue) at the same time.

1. Linear tone mapping in log domain :

2. Tone mapping in gradient domain :

3. Response curve :

p.s.

In sampling step, I sort all pixels in the image with largest exposure. Force sampling the brightest pixel, then sampling the others averagely.

HDR - 6

alpha = 0.09

beta = 0.9

s = 0.1

HDR - 5

Some bugs are fixed. Here is new radiance map :

linear tone mapped in log domain :

Gradient attenuation map :

linear mapping after attenuation in log domain :

LDR with exp(I) :

Final result :

There are still some problems :

1. How to get better result from Poisson equation ?

2. How to select alpha & beta ?

3. Final result is not as good as the one presented in original paper.

HDR - 4

I think ...... I didn`t add any special effect ......

HDR - 3

what`s this ? Should I commit suicide ?

HDR - 2

I have tried to impl gradient domain tone mapping recently since the paper said it is "simple". But I still can`t get the correct result.

Above is the result of Poisson equation. After this, I don`t know how to mapping all channels. Below is one of my test which mixed 3 result (rgb) of poission equation.

Hope I can solve it in a few days /_\.

HDR

Recently, I have to check if the knowledge of HDR in my brain is correct. Here are some result :

1. One of Paul Debevec`s data set. Since my camera "can not" manually change shutter speed and I don`t want to implement MTB currently, I test the algorithm by this data set.

2. Rough tone mapped result :

3. Rough radiance map :

The response curve recovery algorithm is not difficult since you can find matlab version in the end of the paper.

stitching - 2

The left most part is incorrect...don`t know why /_\

Both of the photo set are taken without tripad.

I got more problems now :

• It took too much time in matching. So......maybe I need something like kd tree ? (but I`m lazy)
• I calculate the camera parameters by SBA one by one. But can`t optimize all of them at the same time (after getting rough result ), the image above are stitched by those "rough" result. There may be some bug in my sba usage.
• Multi-band blending ?

stitching

Apply SBA, still can`t get good result and don`t known how to get better focal length.

20100408

p.s.

Now I got better result due to fix a typo in DLT. XD

stitching with DLT

Quick testing with direct linear transformation. The result is better than affine transformation. I guess it`s time to use bundle adjustment.

SIFT with up scale

without up scale :

with up scale :

Up scaling provide about 50% matched features. Now I have to work on memory & cpu usage balance.

Affine Stitching

After adding hat function, clamping descriptor and remap float to u8, I got better result. The paper weighted each bin in descriptor with 1 - d. I don`t know why I can`t get better result. But clamp magnitude to 0.01 ~ 0.99 works fine in my implementation.

I can`t wait to try this ! So I made a simple implementation to stitch image : SIFT + Affine + RANDSAC + SVD. Replace affine model should get better result, I will try it later.

SIFT in real scene

My implementation is still not as robust as imagination. This is one of my better test.

SIFT

I still have many question marks flying around my head. There may be some miss-understanding that make my implementation is weak for scaling & orientation invariance (which should be the strongest compared with the other algorithms).

Ref :

• Rob Hess
• David G. Lowe

My code is heavily base on Rob Hess`s implementation. But I don`t want to install opencv and can`t understand the descriptor building function. So I just sampling a rotated grid with fixed size in all scale. In the matching step, check distance between 2 descriptor only (accepted if d is less than 0.001)

20100223

Get better result after pre-smoothing.

DOG

A rough DOG result between 2 scale level. (float value are clamped between 0.0f - 1.0f)