Ying Cui

N(λ1​ℒfocal​(Mt,Mgtt)+λ2​ℒdice​(Mt,Mgtt))\mathcal{L}=\sum_{t=1}^{N}\left(\lambda_{1}\mathcal{L}_{\text{focal}}(M^{t},M_{\text{gt}}^{t})+\lambda_{2}\mathcal{L}_{\text{dice}}(M^{t},M_{\text{gt}}^{t})\right) (2) where λ1\lambda_{1} and λ2\lambda_{2} balance the contribution of each loss term. Focal Loss helps the model to emphasize hard-to-classify pixels, thereby improving its robustness to small, ambiguous, or partially occluded moving objects. Dice Loss mitigates the foreground-background class imbalance by optimizing the spatial overlap between predictions and ground-truth, thereby achieving more stable and coherent motion prediction. Together with the multi-frame attention in the motion decoder and the, D motion cues such as optical flow [44, 46], struggle to distinguish independent object motion from camera motion due to the lack of depth difference. Moreover, the short temporal perception of optical flow make the models sensitive to occlusion and can not adapt to long-term requirements. To solve the issues, researchers devote to leveraging