Fig 1 - uploaded by Ramanathan Sethuraman
Content may be subject to copyright.
Source publication
In the domain of motion estimation based applica-tions, in order to keep the bandwidth requirements low, the usage of multiple levels of memory hierarchy is a necessity. Usually, the first level (L0) holds the search area of the estimator and the second (L1) holds the region wherein the estimation is performed. This work presents the region-based m...
Contexts in source publication
Context 1
... the high bandwidth requirements. We focus in this work to the motion estimation based video applications like de-interlacing and picture-rate up-conversion. As our starting point, we select the model utilizing the two-level memory hierarchy as shown in figure 1 introduced in [2]. In the mentioned applications, the L0 scratchpad (marked as L0 in fig. 1) holds the so-called Search Area (SA) of the motion estimator/compensator. The L1 scratchpad is introduced in order to reduce the bandwidth requirements towards the off-chip image memory to minimum. The L1 scratchpad can hold one stripe of the image [2], [3] or one region of the image [4], [5], [6]. Both methods have their advantages ...
Context 2
... work [2] analyses the impact of different levels of memory hierarchy to the bandwidth requirements towards the off-chip image memory. In this model, the L0 scratchpad holds the search area of the motion estimator/compensator (the architecture is as depicted in fig. 1). The introduction of stripe-based L1 scratchpad reduced the number of pixel accesses from image memory to only one access per pixel. This is enabled since the stripe-based L1 scratchpad holds SL1 Y (usually equivalent to the height of the SA, L0 Y ) SA Fig. 3. Picture illustrates the search area, SA (size 9*5 blocks) centered around ...
Context 3
... in the case of fixed aspect ratios, in the case of dynamic aspect ratios, the convergence process is finished in the frame number 4. The complete convergence was prevented in the frame number 3 since there was the object that vertically spans across the screen (the girl and her shadow). Further, there is a clear perceivable difference in quality in fig 11 compared to fig 10. Starting from frame number 3, we see that the floor (tiles) looks much clearer in the case of extreme dynamic ratios, since the correct velocity is estimated. ...
