Skip to content

Latest commit

 

History

History
48 lines (35 loc) · 2.16 KB

EVA2_Exploiting Temporal Redundancy in Live Computer Vision.md

File metadata and controls

48 lines (35 loc) · 2.16 KB
Raw

Paper title:

EVA2: Exploiting Temporal Redundancy in Live Computer Vision

Publication:

ISCA’18

Problem to solve:

Video tasks consumes lot of energy and hardware, which needs to be accelerated

Major contribution:

  1. Design the activation motion compensation (AMC) algorithm. AMC captures visual motion in the input video and uses it to transform saved CNN activations. When pixels move in the input scene, AMC moves the corresponding values in the activation data. The algorithm skips a series of layers in the CNN by predicting their output and then invokes the remaining layers to compute the final vision result.

Efficient Motion Estimation: in this paper, they propose a new block matching algorithm to estimate the motion of entire receptive fields. As for how to choose the target layer, this choice controls both AMC’s potential efficiency benefits and its error rate. A later target layer lets AMC skip more computation during predicted frames, but a larger CNN prefix can also compound the influence of layers that make activation warping imperfect. AMC choose the latest convolution layer.

The primary control that AMC has over vision accuracy and execution efficiency is the allocation of key frames, which are both more expensive and more accurate than predicted frames. Several strategies exist to decide when to use each type of frame. The simplest is a static key frame rate: every nth frame is a key frame, and the rest are predicted frames. A adaptive strategy computes the match error for each pixel block in the scene. When the aggregate error across all blocks is high, this strategy allocates a new key frame.

  1. Design a new hardware module, the Embedded Vision Accelerator Accelerator to implement AMC algorithm.

EVA2 is not a complete CNN accelerator. Instead, in this paper, the authors design it to complement existing deep learning accelerators.

Lessons learnt:

Except for frame based acceleration, maybe block based benefits for accuracy. In the paper, the block matching algorithm considers both the redundant computations and the receptive field, which is quite elegant.