Tim Meinhardt, Matt Feiszli, Yuchen Fan, Laura Leal-Taixe, Rakesh Ranjan
Until recently, the Video Instance Segmentation (VIS) community operated under the common belief that offline methods are generally superior to a frame by frame online processing. However, the recent success of online methods questions this belief, in particular, for challenging and long video sequences. We understand this work as a rebuttal of those recent observations and an appeal to the community to focus on dedicated near-online VIS approaches. To support our argument, we present a detailed analysis on different processing paradigms and the new end-to-end trainable NOVIS (Near-Online Video Instance Segmentation) method. Our transformer-based model directly predicts spatio-temporal mask volumes for clips of frames and performs instance tracking between clips via overlap embeddings. NOVIS represents the first near-online VIS approach which avoids any handcrafted tracking heuristics. We outperform all existing VIS methods by large margins and provide new state-of-the-art results on both YouTube-VIS (2019/2021) and the OVIS benchmarks.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Video Instance Segmentation | YouTube-VIS 2021 | AP50 | 82 | NOVIS (Swin-L) |
| Video Instance Segmentation | YouTube-VIS 2021 | AP75 | 66.5 | NOVIS (Swin-L) |
| Video Instance Segmentation | YouTube-VIS 2021 | AR1 | 47.9 | NOVIS (Swin-L) |
| Video Instance Segmentation | YouTube-VIS 2021 | AR10 | 64.4 | NOVIS (Swin-L) |
| Video Instance Segmentation | YouTube-VIS 2021 | mask AP | 59.8 | NOVIS (Swin-L) |
| Video Instance Segmentation | YouTube-VIS 2021 | AP50 | 69.4 | NOVIS (ResNet-50) |
| Video Instance Segmentation | YouTube-VIS 2021 | AP75 | 50 | NOVIS (ResNet-50) |
| Video Instance Segmentation | YouTube-VIS 2021 | AR1 | 41.3 | NOVIS (ResNet-50) |
| Video Instance Segmentation | YouTube-VIS 2021 | AR10 | 54.4 | NOVIS (ResNet-50) |
| Video Instance Segmentation | YouTube-VIS 2021 | mask AP | 47.2 | NOVIS (ResNet-50) |
| Video Instance Segmentation | YouTube-VIS validation | AP50 | 75.7 | NOVIS (ResNet-50) |
| Video Instance Segmentation | YouTube-VIS validation | AP75 | 56.9 | NOVIS (ResNet-50) |
| Video Instance Segmentation | YouTube-VIS validation | AR1 | 50.3 | NOVIS (ResNet-50) |
| Video Instance Segmentation | YouTube-VIS validation | AR10 | 60.6 | NOVIS (ResNet-50) |
| Video Instance Segmentation | YouTube-VIS validation | mask AP | 52.8 | NOVIS (ResNet-50) |
| Video Instance Segmentation | OVIS validation | AP50 | 68.3 | NOVIS (Swin-L) |
| Video Instance Segmentation | OVIS validation | AP75 | 43.8 | NOVIS (Swin-L) |
| Video Instance Segmentation | OVIS validation | AR1 | 19.4 | NOVIS (Swin-L) |
| Video Instance Segmentation | OVIS validation | AR10 | 46.9 | NOVIS (Swin-L) |
| Video Instance Segmentation | OVIS validation | mask AP | 43.5 | NOVIS (Swin-L) |
| Video Instance Segmentation | OVIS validation | AP50 | 56.2 | NOVIS (ResNet-50) |
| Video Instance Segmentation | OVIS validation | AP75 | 32.6 | NOVIS (ResNet-50) |
| Video Instance Segmentation | OVIS validation | AR1 | 15.7 | NOVIS (ResNet-50) |
| Video Instance Segmentation | OVIS validation | AR10 | 37.1 | NOVIS (ResNet-50) |
| Video Instance Segmentation | OVIS validation | mask AP | 32.7 | NOVIS (ResNet-50) |