| Presentation | 2021-09-17 Improving Mask Generation Accuracy Exploiting Optical Flow in Weakly Supervised Instance Segmentation Jun Ikeda, Junichiro Mori, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Weakly supervised instance segmentation is important because it reduces the huge pixel-level annotation cost required to train models. One of the challenges in the weakly supervised approaches which rely on instance-level class labels and bounding boxes is to efficiently learn foreground features by separating the foreground from the background in the bounding box. However, existing approach often misrecognize the background area as the foreground due to the imperfect separation by the local color similarity. We focus on the observation that the foreground is likely to move differently from the backgound, and show that observing the difference in optical flow enable us to separate them in a different way than color. Then, we propose considering the optical flow similarity in addition to the color similarity to generate the pseudo labels for mask head training. Moreover, we demonstrate that our model outperforms the existing method on YouTube-VIS and that our model improves the misrecognition. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | instance segmentation / weakly supervised learning / optical flow / video |
| Paper # | MVE2021-15 |
| Date of Issue | 2021-09-10 (MVE) |
| Conference Information | |
| Committee | MVE |
|---|---|
| Conference Date | 2021/9/17(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Online |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | Masayuki Ihara(RIKEN) |
| Vice Chair | Kiyoshi Kiyokawa(NAIST) |
| Secretary | Kiyoshi Kiyokawa(Oosaka Inst. of Tech.) |
| Assistant | Naoya Isoyama(NAIST) / Takenori Hara(DNP) / Mitsuhiro Goto(NTT) |
| Paper Information | |
| Registration To | Technical Committee on Media Experience and Virtual Environment |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Improving Mask Generation Accuracy Exploiting Optical Flow in Weakly Supervised Instance Segmentation |
| Sub Title (in English) | |
| Keyword(1) | instance segmentation |
| Keyword(2) | weakly supervised learning |
| Keyword(3) | optical flow |
| Keyword(4) | video |
| 1st Author's Name | Jun Ikeda |
| 1st Author's Affiliation | The University of Tokyo(UTokyo) |
| 2nd Author's Name | Junichiro Mori |
| 2nd Author's Affiliation | The University of Tokyo(UTokyo) |
| Date | 2021-09-17 |
| Paper # | MVE2021-15 |
| Volume (vol) | vol.121 |
| Number (no) | MVE-179 |
| Page | pp.pp.38-43(MVE), |
| #Pages | 6 |
| Date of Issue | 2021-09-10 (MVE) |