Jacob Ström

i PACKMAN: high-quality, low-complexity texture compression for mobile phones

We present a novel texture compression scheme, called iPACKMAN, targeted for hardware implementation. In terms of image quality, it outperforms the previous de facto standard texture compression algorithms in the majority of all cases that we have tested. Our new algorithm is an extension of the PACKMAN texture compression system, and while it is a bit more complex than PACKMAN, it is still very low in terms of hardware complexity.

ETC2: texture compression using invalid combinations

We present a novel texture compression system for improved image quality. Building on the iPACKMAN/ETC method, bit combinations that are invalid in that system are used to allow for three additional decompression modes without increasing the bit rate. These modes increase quality, especially for color edges and blocks with smoothly varying content. Due to the use of invalid bit combinations, the system, called ETC2, is backwards compatible with iPACKMAN/ETC. It outperforms S3TC/DXTC and iPACKMAN/ETC in terms of PSNR with 0.8 dB and 1.0 dB respectively, which is clearly visible to the human eye.

PACKMAN: texture compression for mobile phones

We present a new lossy texture compression (TC) scheme, targeted for mobile devices, which compresses 2× 4 pixel blocks into 32 bits. Compared to our POOMA TC work [Akenine-Moller and Strom 2003], peak signal to noise ratio (PSNR) is improved by 0.6 dB on average, the bit rate is reduced from 5.33 bits per pixel (bpp) to 4 bpp, and we avoid 2×3 blocks that are awkward for hardware implementation. Most mobile devices, such as phones and PDAs, have very limited available memory bandwidth, e.g., 16 or 32 bits/cycle. For 3D graphics applications, TC is therefore a fundamental technique for reducing bandwidth usage. Most TC schemes compress each block of pixels individually. In general it is simpler to compress larger blocks efficiently, and therefore the most common block size is 4× 4 pixels compressed into 64 bits (S3TC, DXTC). If a compressed block occupies the same number of bits as the bus width (e.g., 32 bits), then pipeline stalls can be avoided [AkenineMoller and Strom 2003], which simplifies hardware implementation. For more information on different TC schemes, we refer to Fenney’s [2003] previous work section.

Model-based real-time head tracking

This paper treats real-time tracking of a human head using an analysis by synthesis approach. The work is based on the Structure from Motion (SfM) algorithm from Azarbayejani and Pentland (1995). We will analyze the convergence properties of the SfM algorithm for planar objects, and extend it to handle new points. The extended algorithm is then used for head tracking. The system tracks feature points in the image using a texture mapped three-dimensional model of the head. The texture is updated adaptively so that points in the ear region can be tracked when the user′s head is rotated far, allowing out-of-plane rotation of up to without losing track. The covariance of the- and the-coordinates are estimated and forwarded to the Kalman filter, making the tracker robust to occlusion. The system automatically detects tracking failure and reinitializes the algorithm using information gathered in the original initialization process.

High quality normal map compression

Normal mapping is a widely used technique in real-time graphics, but so far little research has focused on compressing normal maps. Therefore, we present several simple techniques that improve the quality of ATIs 3Dc normal map compression algorithm. We use varying point distributions, rotation, and differential encoding. On average, this improves the peak-signal-to-noise-ratio by 3 dB, which is clearly visible in rendered images. Our algorithm also allows us to better handle slowly varying normals, which often occurs in real-world normal maps. We also describe the decoding process in detail.

Floating-point buffer compression in a unified codec architecture

This paper presents what we believe are the first (public) algorithms for floating-point (fp) color and fp depth buffer compression. The depth codec is also available in an integer version. The codecs are harmonized, meaning that they share basic technology, making it easier to use the same hardware unit for both types of compression. We further suggest to use these codecs in a unified codec architecture, meaning that compression/decompression units previously only used for color-and depth buffer compression can be used also during texture accesses. Finally, we investigate the bandwidth implication of using this in a unified cache architecture. The proposed fp16 color buffer codec compresses data down to 40% of the original, and the fp16 depth codec allows compression down to 4.5 bpp, compared to 5.3 for the state-of-the-art int24 depth compression method. If used in a unified codec and cache architecture, bandwidth reductions of about 50% are possible, which is significant.

Removing motion blur from barcode images

Camera shake during photography is a common problem which causes images to get blurred. Here we choose a specific problem in which the image is a barcode and the motion can be modeled as a convolution. We design a blind deconvolution algorithm to remove the translatory motion from a blurred barcode image. Based on the bimodal characteristics of barcode image histograms, we construct a simple target function that measures how similar a deconvoluted image is to a barcode. We minimize this target function over the set of possible convolution kernels to find the most likely blurring kernel. By restricting our search to dome-shaped kernels (first monotonously increasing and then monotonously decreasing) we decrease the number of false solutions. We have tried our system on a collection of a 138 barcode images with varying camera blur, and the recognition rate increases from 32% to 65%.

Table-based Alpha Compression

In this paper we investigate low‐bitrate compression of scalar textures such as alpha maps, down to one or two bits per pixel. We present two new techniques for 4 × 4 blocks, based on the idea from ETC to use index tables. We demonstrate that although the visual quality of the alpha maps is greatly reduced at these low bit rates, the quality of the final rendered images appears to be sufficient for a wide range of applications, thus allowing bandwidth savings of up to 75%. The 2 bpp version improves PSNR with over 2 dB compared to BTC at the same bit rate. The 1 bpp version is, to the best of our knowledge, the first public 1 bpp texture compression algorithm, which makes comparison hard. However, compared to just DXT5‐compressing a subsampled texture, our 1 bpp technique improves PSNR with over 2 dB. Finally, we show that some aspects of the presented algorithms are also useful for the more common bit rate of four bits per pixel, achieving PSNR scores around 1 dB better than DXT5, over a set of test images.

Lossless compression of already compressed textures

Texture compression helps rendering by reducing the footprint in graphics memory, thus allowing for more textures, and by lowering the number of memory accesses between the graphics processor and memory, increasing performance and lowering power consumption. Compared to image compression methods like JPEG however, textures codecs are typically much less efficient, which is a problem when downloading the texture over a network or reading it from disk. Therefore, in this paper we investigate lossless compression of already compressed textures. By predicting compression parameters in the image domain instead of in the parameter domain, a more efficient representation is obtained compared to using general compression such as ZIP or LZMA. This works well also for pixel indices that have previously proved hard to compress. A 4-bit-per-pixel format can thus be compressed to around 2.3 bits per pixel (bpp), or 9.6% of the original size, compared to around 3.0 bpp when using ZIP or 2.8 bpp using LZMA. Compressing the original images with JPEG to the same quality also gives 2.3 bpp, meaning that texture compression followed by our packing is on par with JPEG in terms of compression efficiency.

Error-bounded lossy compression of floating-point color buffers using quadtree decomposition

In this paper, we present a new color buffer compression algorithm for floating-point buffers. It can operate in either an approximate (lossy) mode or in an exact (lossless) mode. The approximate mode is error-bounded and the amount of introduced accumulated error is controlled via a few parameters. The core of the algorithm lies in an efficient representation and color space transform, followed by a hierarchical quadtree decomposition, and then hierarchical prediction and Golomb–Rice encoding. We believe this is the first lossy compression algorithm for floating-point buffers, and our results indicate significantly reduced color buffer bandwidths and negligible visible artifacts.