- 1、本文档共14页,可阅读全部内容。
- 2、原创力文档(book118)网站文档一经付费(服务费),不意味着购买了该文档的版权,仅供个人/单位学习、研究之用,不得用于商业用途,未经授权,严禁复制、发行、汇编、翻译或者网络传播等,侵权必究。
- 3、本站所有内容均由合作方或网友上传,本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺!文档内容仅供研究参考,付费前请自行鉴别。如您付费,意味着您自己接受本站规则且自行承担风险,本站不退款、不进行额外附加服务;查看《如何避免下载的几个坑》。如果您已付费下载过本站文档,您可以点击 这里二次下载。
- 4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等,请点击“版权申诉”(推荐),也可以打举报电话:400-050-0827(电话支持时间:9:00-18:30)。
查看更多
Instance Embedding Transfer to Unsupervised Video Object Segmentation
8
1
0 Abstract
2
b We propose a method for unsupervised video object
e segmentation by transferring the knowledge encapsulated
F in image-based instance embedding networks. The in-
7 stance embedding network produces an embedding vector
2 for each pixel that enables identifying all pixels belong-
] ing to the same object. Though trained on static images, Figure 1. An example of the changing segmentation target (fore-
V the instance embeddings are stable over consecutive video ground) in videos depending on motion. A car is the foreground
C frames, which allows us to link objects together over time. in the top video while a car is the background in the bottom video.
. Thus, we adapt the instance networks trained on static im- To address this issue, our method obtains embeddings for ob-
s ages to video object segmentation and incorporate the em- ject instances and identifies representative embeddings for fore-
c
[ beddings with objectness and optical flow features, with- ground/background and then segments the frame based on the rep-
out model retraining or online fine-tuning. The proposed resentative embeddings. Left: the ground truth. Middle: A visu-
2 method outperforms state-of-the-art unsupervised segmen- alization of the embeddings projected into RGB space via PCA,
v tation methods in the DAVIS dataset and the FBMS dataset. along with representative points for the foreground (magenta) and
8 background (blue). Right: the segmentation masks produced by
0 the propos
文档评论(0)