Computation Error Tolerance in Motion Estimation Algorithms

Abstract

In this paper we study the computation error tolerance properties of motion estimation algorithms. We are motivated by two scenarios where hardware systems may introduce computation errors. First, we consider hardware faults such as those arising in a typical fabrication process. Second, we consider "soft" errors due to voltage scaling, which can arise when operating at a lower voltage than specified for the system. Current practice is to discard all faulty systems. However there is an increasing interest in tools that can identify faulty systems which provide acceptable performance. We show that motion estimation (ME) algorithms exhibit significant error tolerance in these two scenarios. We propose simple error models and use these to provide insights into what features in these ME algorithms lead to increased error tolerance. Our comparison of the full search ME and a state of the art fast ME approach in the context of H.264/AVC shows that while both techniques are error tolerant, the faster algorithm is in fact more robust to computation errors.

Power Efficient Motion Estimation using Multiple Imprecise Metric Computations

“…Our previous work has demonstrated that image/video compression systems exhibit ET characteristics, even if no explicit error control block is added in the presence of VOS [2] (additional work dealing with hard errors due to deterministic faults shows similar results [7,5]). For example, our simulations demonstrated 37% power savings in the ME process with negligible performance penalty in typical video encoders.…”

“…Note that a 20% rate increase would correspond to around 0.5dB PSNR loss if the rate were to be kept fixed rate, while 5% rate increase leads to less than 0.1dB loss [2]. Also in [18], three step search (TSS) algorithm was used, which is shown to be less resilient to soft errors than EPZS, thus rate increase can be worse if we use TSS instead of EPZS [2]. In the extreme case, when we use FS algorithm, MAX estimator shows similar rate increase, but threshold estimator shows more than 90% increase, which corresponds to more than 2dB loss.…”

“…In the other range of parameters all new estimators show reasonable performance; with all estimators leading to less than 5% rate increase. Note that a 20% rate increase would correspond to around 0.5dB PSNR loss if the rate were to be kept fixed rate, while 5% rate increase leads to less than 0.1dB loss [2]. Also in [18], three step search (TSS) algorithm was used, which is shown to be less resilient to soft errors than EPZS, thus rate increase can be worse if we use TSS instead of EPZS [2].…”

Dynamic Voltage Scaling Algorithms for Power Constrained Motion Estimation

“…This performance penalty may be acceptable for specific application scenarios. We have also provided a primarily experimental evaluation of the behavior of several ME algorithms under "soft error" conditions applying soft DSP approaches to MMC within ME process [3].…”

“…In our previous work, we have shown that image/video compression systems exhibit error tolerance (ET) characteristics, even if no explicit error control block is added, and this under both hard errors (due to deterministic faults) [6,5] and soft errors (due to DVS) [3]. Errors due to VOS in these applications are either i) concealed by other parts of the system (e.g., quantization can conceal errors affecting a transform computation) or ii) are "acceptable" [2].…”