Abstract
In this work we present a novel solution for Video Instance Segmentation (VIS), that is automatically generating instance level segmentation masks along with object class and tracking them in a video. Our method improves the masks from segmentation and propagation branches in an online manner using the Mask Selection Network (MSN) hence limiting the noise accumulation during mask tracking. We propose an effective design of MSN by using patch-based convolutional neural network. The network is able to distinguish between very subtle differences between the masks and choose the better masks out of the associated masks accurately. Further, we make use of temporal consistency and process the video sequences in both forward and reverse manner as a post processing step to recover lost objects. The proposed method can be used to adapt any video object segmentation method for the task of VIS. Our method achieves a score of 49.1 mAP on 2021 YouTube-VIS Challenge and was ranked third place among more than 30 global teams.
Introduction
Video Instance Segmentation (VIS) is a challenging computer vision task that requires simultaneous detection, segmentation, and tracking of object instances across video sequences. Unlike static image segmentation, VIS must maintain identity consistency over time, often in the face of occlusions, motion blur, and rapid appearance changes. A primary bottleneck in existing VIS frameworks is noise accumulation: errors in mask propagation from one frame can propagate and amplify in subsequent frames, leading to significant "drift." Furthermore, VIS benchmarks heavily penalize misclassifications, making robust object identification as critical as mask quality.
To address these issues, we propose the Mask Selection Network (MSN). MSN acts as an intelligent "quality control" gate that evaluates mask proposals from different sources—specifically, new detections and propagated tracks—and selects the highest quality candidate. This approach effectively "resets" the error accumulation at each frame, ensuring long-term tracking stability.
Methodology
Our framework adopts an online processing strategy composed of three main components: instance segmentation, mask propagation, and selection.
Instance Segmentation and Propagation Branches
The system utilizes a high-performance instance segmentation branch featuring a Swin Transformer backbone and a Cascade Mask R-CNN head. This branch provides robust initial detections and classifications. In parallel, a propagation branch employs the Collaborative Video Object Segmentation by Foreground-Background Integration (CFBI) method to track masks from the previous frame into the current one.
Mask Selection Network (MSN)
The MSN is the core innovation of our work. It is designed to resolve discrepancies between the segmentation and propagation branches. For any given object instance, the network receives two candidate masks. MSN is implemented as a patch-based convolutional neural network. By analyzing local patches rather than global image features, MSN can identify subtle boundary errors and fine-grained quality differences that global networks might overlook.
As shown in the architecture diagram, the MSN takes an 8-channel input (comprising the image and mask information) and outputs a selection score. This design is highly efficient; it achieves a selection accuracy of 74% while utilizing only 2.22 GMacs, representing a nearly 9x reduction in FLOPs compared to previous state-of-the-art selectors (19.45 GMacs).
Temporal Post-Processing
To further enhance temporal consistency and handle long-term occlusions, we implement a bidirectional processing strategy. Video sequences are processed in both forward and reverse directions. The results from both passes are associated using a Hungarian algorithm, allowing the system to recover tracks for objects that may have been missed or poorly segmented in a single-pass forward tracking scenario.
Experimental Results
The proposed method was rigorously evaluated on the YouTube-VIS 2021 dataset, a benchmark containing 40 object classes.
Challenge Performance
Our approach achieved a mean Average Precision (mAP) of 49.1 in the 2021 YouTube-VIS Challenge, placing 3rd globally among more than 30 competing teams. This result demonstrates the effectiveness of the MSN-based pipeline in complex, real-world video scenarios.
Ablation Analysis
Ablation studies confirmed that the Mask Selection Network is the most significant contributor to performance gains, providing a +2.6 mAP boost. This indicates that the ability to intelligently choose between detection and propagation is more impactful than standard techniques like multi-scale testing. Additionally, our model showed remarkable consistency at high Intersection-over-Union (IoU) thresholds, signifying superior mask boundary quality.
Conclusion
We have presented MSN, an efficient and effective network for mask selection in Video Instance Segmentation. By integrating a patch-based CNN into the tracking pipeline, we successfully mitigate noise accumulation and improve mask accuracy. While the current multi-stage pipeline achieves state-of-the-art results, future work will explore end-to-end integration to reduce deployment complexity and further optimize classification performance.
Summary updated