LiVOS: Light Video Object Segmentation with Gated Linear Matching

We introduced four primary baselines in the main paper. In this section, we present the remaining seven baselines. AOT [24] and DeAOT [22] are two consecutive approaches to improve the efficiency of VOS with multiple objects. Following Cutie [6], we use the model variants with a ResNet-50 backbone as baselines. CFBI [23] and CFBI+ [25] propose a collaborate VOS approach that integrates both foreground and background information into embedding learning. As they only use two memory frames, we classify them as non-STM methods with less strict criteria. Both models use RestNet-101 as the backbone, and we adopt them as our baselines. DEVA [5] decouples task-specific image-level segmentation and mask propagation for universal video segmentation. We use as the model trained solely on YouTube-VOS [21] and DAVIS 2017 [16] as the baseline. SwiftNet [19] balances accuracy and speed by compressing spatiotemporal redundancy in matching-based VOS with a pixel-adaptive memory. We use the model variant with a ResNet-50 backbone as the baseline. MobileVOS [15] distills knowledge from a teacher model utilizing large backbone and infinite memory. We use the best-performing model variant with a ResNet-18 backbone as the baseline.