International organisation for standardisation organisation internationale de normalisation

4TE2: IBDI and memory compression

4.1Summary

4.1.1.1.1.1.1.1.1JCTVC-C076 TE2: Summary of TE2 on IBDI and memory compression [T. Chujoh (TE coordinator)]

This contribution is a summary of activies and results relating to tool experiment 2: IBDI and memory compression as described in JCTVC-B302r2. Nine companies and one university had been registered in TE2 and four proposed tools had been evaluated on the common condition with cross-verification. There are two test conditions. One is the case of keeping the bit depth of reference memory to the input bit depth on IBDI and the other is the case of reduction of reference memory access bandwidth and size.

A substantial amount of information was in this summary report.

The proposed techniques were summarized as shown in the table below:

Tool	Proponent	Cross-checker	Method
1	NEC	Panasonic, JVC	DPCM of 8 samples
2	Panasonic	TOSHIBA	Hadamard transform and bit-plane coding
3	TI	TOSHIBA	Wavelet transform and DC prediction by 8x8 block
4	TOSHIBA	NEC, Yonsei Unv.	Adaptive scaling by 4x4 block and de-scaling by sample

The summary report included tables illustrating the impact on measures relating to coding efficiency, complexity, and memory access for HE and LC configurations

It was remarked that for the various memory compression schemes in general, the behavior on static regions may be a particular concern.

It was also noted that more complex techniques need to be significantly better than 8 bit rounding of IBDI results in order to to be useful.

4.21-D DPCM-based memory compression

4.2.1.1.1.1.1.1.1JCTVC-C093 TE2: 1-D DPCM-based memory compression [H. Aoki, K. Chono, K. Senzaki, J. Tajime, Y. Senda (NEC)]

In this contribution, a performance report on the 1-D DPCM-based memory compression method described in JCTVC-B057 was presented. The method was designed to reduce reference frame memory size and memory access bandwidth for motion compensation, with relatively low complexity. The method employs 1-D DPCM. Experimental results reportedly showed that average coding losses are 10.7% for test cases without IBDI and 2.0% for test cases with 12-bit IBDI. Subjective quality degradation caused by such coding loss was reported to be invisible for most test cases. Average increase of decoding time is as small as 3.0% and memory bandwidth reduction ratios are 40.6% for test cases without IBDI and 49.0% for test cases with 12-bit IBDI. Experimental results on bandwidth reportedly showed that bandwidth reduction depends on memory architecture and implementation on mapping of compressed image data onto frame memory.

It was remarked that the 2% loss with IBDI seems to be most of the gain that IBDI provides.

It was remarked that such a proposal should be compared against just rounding to 8 bits when storing.

JCTVC-C094 and JCTVC-C095 are related proposal documents for a new proposal.

It was asked whether the decoder is required to do clipping after reconstruction of each sample. The proponent indicated that this is not necessary. Instead, the encoder must perform quantization in a way that avoids reconstruction overflow.

4.2.1.1.1.1.1.1.2JCTVC-C074 TE2: Cross-check of memory compression results from NEC (JCTVC-B057/JCTVC-C093) [C. S. Lim, V. Wahadaniah, S. Naing, H. W. Sun (Panasonic)] (had been uploaded on the ITU ftp site and was only later put on the new site)

The purpose of this document was to present a cross-check of memory compression results from NEC (JCTVC-B057/JCTVC-C093) for Tool Experiment 2 (IBDI and memory compression). The cross-check was reportedly completed successfully and the results from NEC were reportedly reproduced without any problem.

4.3Reference frame compression using image coder

4.3.1.1.1.1.1.1.1JCTVC-C073 TE2: Reference frame compression using image coder [C. S. Lim, V. Wahadaniah, S. Naing, H. W. Sun (Panasonic)]

This contribution presented results of Tool Experiment 2 (IBDI and memory compression) from the reference frame compression scheme proposed by Panasonic (JCTVC-B103). The presented reference frame compression scheme uses an image coder comprising a transform, scanning, and bit-plane coding. The presented scheme was implemented on TMuC version 0.7.3 revision 188 and experiments were conducted using the common test conditions defined in the Tool Experiment 2 document (JCTVC-B302).

The memory compression rate used in the experiments for High Efficiency setting was 12-bit to 5.33-bit (44.4%). In comparison with the TMuC software without reference frame compression, experiment results reportedly show an average coding efficiency drop of 2.9% for the High Efficiency, Random Access setting and an average coding efficiency drop of 4.7% for the High Efficiency, Low Delay setting.

The memory compression rate used in the experiments for Low Complexity setting was 8-bit to 5.33-bit (66.7%). In comparison with the TMuC software without reference frame compression, experiment results reportedly show an average coding efficiency drop of 2.9% for the Low Complexity, Random Access setting and an average coding efficiency drop of 4.0% for the Low Complexity, Low Delay setting. It was noted from the results in all test settings that the larger drop in performance occurs in the lower resolution images.

It was remarked that the complexity of the proposed reference frame compression method seems relatively high, and the coding efficiency impact is not negligible.

4.3.1.1.1.1.1.1.2JCTVC-C178 TE2: Cross-check of memory compression results from Panasonic (JCTVC-C074) [A. Segall (Sharp)] (missing prior, available first day)

This contribution provided a cross-check of memory compression as proposed by NEC and results provided by Panasonic. This cross-check was performed within the context of TE2. The experimental results reportedly closely matched the rate-distortion results provided by Panasonic. The reported maximum Y BD Rate differed from the provided results by at most 0.1%. The cross-checker indicated that he had read through the software and found it to match the proposal algorithm.