In recent years, the live-streaming industry has grown significantly and become a crucial form of online interaction and marketing. Douyin has emerged as one of the largest information streaming platforms. Personalized recommendation, effectively matching users' interests with relevant content, is essential for the prosperity of live streaming and short videos. Unlike short-video recommendations, live-streaming necessitates a bilateral optimization framework that jointly considers real-time content evolution and mutual influence between streamers and audiences. However, existing approaches predominantly focus on static user-item modeling or unilateral consumer-side metrics, failing to capture the spatiotemporal dependencies in multi-behavior traffic or address the interdependence between production and consumption. To bridge these gaps, we propose a novel paradigm for live-streaming recommendation from a bilateral perspective. First, we design Self Flow, a real-time data-streaming engine that generates multi-behavior traffic sequences at minute-level granularity, redefining traffic efficiency metrics to align streamer incentives with audience engagement. Leveraging this infrastructure, we present the Time-Slice level Spatial-Temporal Fusion Network (TSSTFN), which integrates spatial-temporal fusion module to model cross-behavior dependencies and temporal evolution patterns simultaneously. Crucially, we design deployment strategies from a bilateral perspective to optimize the experiences of both the consumption side and the supply side simultaneously. Extensive experiments on Douyin platforms demonstrate significant bilateral improvements: +1.173%/+0.721% watch duration (Douyin/Douyin Lite) for audiences and + 1.037% live duration for streamers. Our fully-deployed solution proves that bilateral value alignment fosters sustainable growth for both content creators and consumers.

Although state-of-the-art methods for human pose estimation have achieved superior results on the single image, their performance on videos usually deteriorates dramatically due to motion blur and occlusion. Since there is close temporal correlation among video frames, exploiting the contextual information properly can be helpful to tackle the problem. In this paper, we present a Temporal Feature Enhancing Network (TFEN) for video human pose estimation. It boosts the per-frame features by utilizing motion information in terms of optical flow and conducting temporal feature encoding by the convolution gated recurrent units (convGRU). It is an end-to-end learning framework and can extend any image based algorithm to video pose estimation. The experimental results validate the effectiveness of the proposed approach on two large-scale video pose estimation benchmarks.