Object-based video coding challenges MPEG

Object-based video coding challenges MPEG
By Junko Yoshida, EE Times
March 22, 2002 (8:51 p.m. EST)
URL: http://www.eetimes.com/story/OEG20020322S0105

PARIS — Has MPEG had its day? A Milpitas, Calif. startup will argue in the affirmative Monday (March 25) when it announces an object-based video coding algorithm that it calls a radical departure from the block-based coding used in all MPEG standards. Scheduled for private previews in Las Vegas next month at the National Association of Broadcasters' (NAB) convention, Pulsent Corp.'s technology — and a host of similar developments both inside and outside the Moving Picture Experts Group — could ignite debate as the industry looks beyond the current slate of standards for digital video compression.

Indeed, Pulsent will hardly be alone in showing improved compression algorithms at NAB. The convention will highlight advances in a host of proprietary and standards-based schemes from such entities as Apple Computer, DemoGraFX, DiVX, iVast, Microsoft Corp., On2 (formerly Duck Corp.) and RealNetworks. Like the Pulsent algorithm, many of the evolving coding technologies — including block-based H.26L, elements of which are being incorporated into a next-generation variant of MPEG-4 — target broadcast-quality video at 1.5 Mbits/second or slower.

At issue is whether the market will be willing to abandon some current infrastructure investments for the sake of increased coding efficiency.

Pulsent believes "the next huge jump in coding efficiency must come from a new discipline," even if it defies conventional wisdom and the existing infrastructure, said chief executive officer Adityo Prakash. "MPEG has done a good job squeezing more out of the existing technology, but we feel that block-based video coding has reached its limits." Rather than rely on block-based motion compensation to predict the next frame, Pulsent's scheme uses an object in one frame to model itself in subsequent frames.

But object-based coding has its detractors. "Not all pictures can be segmented into objects," said Didier LeGall, vice president of home med ia products at LSI Logic Corp. (Milpitas). "To do object-based video coding for a particular class of a picture is possible, but to do it for generic materials — sports, news, movies, science programs and the like — is very difficult." LeGall professed to have more "skepticism than faith" that such schemes will prove practical.

Mature, not old

Others take umbrage at Pulsent's claim that MPEG is running out of gas. "It's not that block-based schemes are at the end of life; it's that schemes that were immature a few years ago have grown up to some degree," said Peter Schirling, delegation head of the U.S. MPEG Committee. "There is still a lot of life left in block-based coding. The investment in such schemes makes them very cost-competitive, and they still deliver quality."

The MPEG community's apparent ace in the hole is H.26L. Developed by the International Telecommunication Union (ITU), H.26L uses block-based coding but includes a number of tools that allow it to deliver muc h more efficient video coding in low bit-rate applications than any current MPEG standard. Moving away from the limitations of MPEG-2's 8 x 8 block discrete cosine transform (DCT) approach, H.26L provides more choices in coding techniques, including multiple reference frames for motion compensation and residual coding based on 4 x 4 blocks and integer transforms.

Change the world

MPEG personnel have joined forces with ITU video coding experts to form the Joint Video Team (JVT), which will pursue an H.26L-based MPEG-4 variant, to be called MPEG-4 part 10, Schirling said.

Describing JVT as "a final, definitive merging of the ITU and MPEG IEC/ISO process," one digital video expert who asked not to be identified predicted that, "the result of the JVT process is going to change the world of MPEG-2 and MPEG-4 forever. Therefore, it will change the world of anything having to do with video forever." By cherry-picking some compression tools designed for H.26L, the source said, "we know we can improve coding efficiencies by 50 percent and achieve resolutions ranging from SDTV to HDTV and digital cinema."

Part 10 is expected to be completed by year's end.

Pulsent, meanwhile, is focusing on what its object-based scheme is capable of doing today. "We can demonstrate full-screen, broadcast-quality video delivered into the home over standard 1.5-Mbit/s DSL lines," Prakash claimed. "We can offer VHS-quality video at 384 kbits/s or less."

"Our algorithm has been extensively tested on a very wide range of video materials, such as an MPEG-2 test sequence suite and other highly demanding video footage provided by various potential partners and customers," Prakash said. Pulsent plans to finish the design of a low-power ASIC for its full codec (both encoder and decoder) by the fourth quarter. Prakash said the startup worked with Japanese consumer companies on its approach but declined to identify them.

Pulsent is targeting low-bandwidth consumer applications such as gateway products for digit al subscriber line (DSL), personal video recorders and possibly HD-DVD. The startup, which ultimately wants to be in the business of intellectual-property core licensing, believes the quickest route to that goal is to prove the viability of its algorithm in its own hardware.

"The architecture is done; major modules of the design are also completed. Right now, we are getting partners' final feedback on the chip spec," Prakash said.

Block-based MPEG-4 does use an object-based technology to add interactivity to its video stream. "But that should not be confused with Pulsent's object-based video coding"' Prakash cautioned. "For MPEG-4, object-based technology was very much an afterthought."

At DemoGraFX, which will also show a coding scheme at NAB, chief architect Tom McMahon said his company's algorithm is based on DCT but at higher precision than any MPEG version. "We will have some startling demonstrations at NAB in the 1-Mbit range," he said.

Ready market?

Some industry obser vers question whether the market will accept Pulsent's radical departure from the status quo. Michelle Abraham, senior analyst for multimedia at Cahners In-Stat Group, had a slew of questions about the approach: "Who else is going to supply chips besides Pulsent? What will be the cost of the customer-premises equipment [for DSL]? What are the license fees for the proprietary technology"

Another question mark is infrastructure. "MPEG-2 is firmly entrenched in digital TV among content owners, service operators and consumer electronics manufacturers," said Matthew Hatch, director of the set-top division at STMicroelectronics. "Our view, and our partners' view, is that there is a long life ahead for MPEG-2."

"In the cable and satellite space," In-Stat's Abraham said, "there is an investment in legacy installed equipment. Echostar and DirecTV would have to switch out 30 million-plus boxes" to switch to a scheme like Pulsent's.

But the startup countered that it intends to coexist with rather than rep lace current industry standards. Pulsent technology's video payloads are designed for compatible transmission through standard MPEG transport stream protocols, said Prakash. "Our technology completely works under the current MPEG-2 transport-based network and infrastructures."

For network services such as video-over-ADSL, "the Pulsent technology can re-encode the digital video from MPEG-2 running at 3 Mbits/s down to a Pulsent stream, running at about 1.5 Mbits/s, which means a single twisted-pair ADSL telephone line could support up to three separate TV sets," said Gerry Kaufhold, principal analyst at In-Stat. But each TV would still need its own set-top box, with a chip capable of decoding Pulsent's algorithm, to decode the video.

The need for transcoding also raised flags. "Transcoding always comes with quality issues," said Richard Bramley, director of R&D for the VLIW-based ST200 program at STMicroelectronics. "Certain know-how would allow you to get around the problem, but transcoding betwe en two very different coding technologies can only make things worse."

Schirling agreed. "In the digital domain we use a lossy scheme," he said, "so decoding, even partial decoding, may cause loss of quality in the final decompression step."

Perhaps the biggest catch is the extensive computation that Pulsent's technology requires for encoding and transcoding. Even Prakash acknowledged that "you can't do a great compression without a lot of computation."

Real-time transcoding

Pulsent's algorithm currently runs in software using a parallel architecture of CPUs. The startup said its ASIC accomplishes transcoding in real-time. A transcoder box sitting in a commercial headend would require multiple ASICs, although gateway products would need only one ASIC, Prakash said.

The company's full-codec ASIC will have between 6 million and 6.5 million gates. An ASIC with a decoding capability only would require less than 2 million gates. Prakash said the ASIC will incorporate "embedded pro cessors along with co-processor engines." Pulsent is licensing an unidentified "basic RISC engine" to make the chip programmable enough to handle such other codecs as MPEG-2 and MPEG-4. The company developed its own DSP-like functionality for the co-processor engines embedded in the ASIC.

Prakash said Pulsent had shopped around for an existing media processor that might run the Pulsent algorithm but found that none was up to the task, "at least not in the current form."

Pulsent will design the front end and hand over its netlist to a service company operated by an EDA vendor to execute the back end. The chip will be fabricated on a 0.13-micron process.