Lossy Event Compression Based On Image-Derived Quad Trees And Poisson Disk Sampling

Abstract

Event cameras have provided new opportunities for tackling visual tasks under challenging scenarios over conventional RGB cameras. However, not much focus has been given on event compression algorithms. The main challenge for compressing events is its unique asynchronous form. To address this problem, we propose a novel event compression algorithm based on a quad tree (QT) segmentation map derived from the adjacent intensity images. The QT informs 2D spatial priority within the 3D space-time volume. In the event encoding step, events are first aggregated over time to form polarity-based event histograms. The histograms are then variably sampled via Poisson Disk Sampling prioritized by the QT based segmentation map. Next, differential encoding and run length encoding are employed for encoding the spatial and polarity information of the sampled events, respectively, followed by Huffman encoding to produce the final encoded events. Our algorithm achieves greater than 6× higher compression compared to the state of the art.

hARMS: A Hardware Acceleration Architecture for Real-Time Event-Based Optical Flow

“…Another method for lossy event-stream compression is the use of quadtrees and Poisson disk sampling as introduced in [14]. This method achieved a compression ratio of more than 6× those presented in prior works.…”

“…This method achieved a compression ratio of more than 6× those presented in prior works. Although this work achieves a high compression ratio, it is limited by its dependence on the use of intensity images generated by an event-based vision sensor such as the DAVIS or RGB-DAVIS [14]. This limitation is a significant drawback as it cannot be used with purely asynchronous temporal-contrast sensors, but instead has to rely on additional intensity data to select the low-priority regions where events can be removed without impacting the quality of event reconstruction.…”

“…This requires an additional three bytes to be transferred for each event containing flow. Therefore, flow events are 11 bytes in total while normal AER events are eight bytes as used in [4], [14]. The value of N nf equals the number of events in the event stream for which no flow can be calculated.…”

“…Some accumulate event frames and use standard image similarity metrics such as structural similarity index measure (SSIM) and peak signal-tonoise ratio (PSNR). These metrics are used in [14], however, they acknowledge that this method gives little insight into temporal errors in the reconstructed event stream. In fact, PSNR and SSIM have substantial limitations even for accumulated event frames.…”

Evaluation of the impact of lossy compression on event camera-based computer vision tasks

“…Besides, the number of events for each signaled position and their EF index are also encoded together by conventional video coding. Another compression method was introduced 33 where, at the encoding stage, events are first aggregated over time to form polarity-based event histograms. A quadtree (QT) segmentation map derived from the adjacent intensity images is then used to solve the lack of spatial structure of event data.…”