Brief

Researchers have developed FoundCAC, a new framework for blind lens aberration correction that utilizes large-scale pre-training and discrete degradation priors. The system improves data scalability by constructing a diverse lens library called AODLibpro and employs a multi-stage vector-quantized representation learning scheme to encode Point Spread Functions into a discrete prior. This approach enables state-of-the-art zero-shot generalization and efficient few-shot adaptation for correcting optical degradations in both synthetic and real-world lenses. AI

Researchers have introduced LFX, a novel unified framework designed to handle various light field (LF) representations for dense semantic segmentation and salient object detection. This framework utilizes a representation-invariant feature modulation space and a Field-of-Parallax Angular Subspace Modeling (FoP-ASM) technique to adapt to different LF data. LFX demonstrates state-of-the-art performance across multiple benchmarks, outperforming specialized methods by significant margins and achieving improved accuracy in both segmentation and detection tasks. AI

Researchers have introduced OmniTrack++, an advanced framework for omnidirectional multi-object tracking designed to overcome challenges like panoramic distortion and identity ambiguity. The system utilizes a feedback-driven approach, incorporating a DynamicSSM block for feature stabilization and FlexiTrack Instances for precise localization and association. To enhance long-term tracking, an ExpertTrack Memory consolidates appearance cues, while a Tracklet Management module adaptively switches between tracking modes based on scene dynamics. The team also released the EmboTrack benchmark, featuring new datasets like QuadTrack and BipTrack, to facilitate evaluation in real-world panoramic scenarios. AI