LVOS: A Benchmark for Long-term Video Object Segmentation

• DOI: 10.48550/arxiv.2211.10181
• Publication Date: 2022-01-01
• Authors: Lingyi Hong; Wenchao Chen; Liu Zhong-ying; Wei Zhang; Pinxue Guo; Zhaoyu Chen; Wenqiang Zhang
• Keywords: Benchmark (surveying); Code (set theory)

Existing video object segmentation (VOS) benchmarks focus on short-term videos which just last about 3-5 seconds and where objects are visible most of the time. These poorly representative practical applications, absence long-term datasets restricts further investigation VOS application in realistic scenarios. So, this paper, we present a new benchmark dataset named \textbf{LVOS}, consists 220 with total duration 421 minutes. To best our knowledge, LVOS is first densely annotated dataset. The 1.59 minutes average, 20 times longer than existing datasets. Each includes various attributes, especially challenges deriving from wild, such as reappearing cross-temporal similar objeccts.Based LVOS, assess algorithms propose Diverse Dynamic Memory network (DDMemory) that three complementary memory banks to exploit temporal information adequately. experimental results demonstrate strength weaknesses prior methods, pointing promising directions for study. Data code available at https://lingyihongfd.github.io/lvos.github.io/.