An important paradigm in 3D object detection is the use of multiple modalities to enhance accuracy in both normal and challenging conditions, particularly for long-tail scenarios. To address this, recent studies have explored two directions of adaptive approaches: MoE-based adaptive fusion, which struggles with uncertainties arising from distinct object configurations, and late fusion for output-level adaptive fusion, which relies on separate detection pipelines and limits comprehensive understanding. In this work, we introduce Cocoon, an object- and feature-level uncertainty-aware fusion framework. The key innovation lies in uncertainty quantification for heterogeneous representations, enabling fair comparison across modalities through the introduction of a feature aligner and a learnable surrogate ground truth, termed feature impression. We also define a training objective to ensure that their relationship provides a valid metric for uncertainty quantification. Cocoon consistently outperforms existing static and adaptive methods in both normal and challenging conditions, including those with natural and artificial corruptions. Furthermore, we show the validity and efficacy of our uncertainty metric across diverse datasets.

Collaborative Perception enables multiple agents, such as autonomous vehicles and infrastructure, to share sensor data via vehicular networks so that each agent gains an extended sensing range and better perception quality. Despite its promising benefits, realizing the full potential of such systems faces significant challenges due to inherent imperfections in underlying system layers, consisting of network layer imperfections and hardware-level noises. Such imperfections and noises include packet loss in vehicular networks, localization errors from GPS measurements, and synchronization errors caused by clock deviation and network latency. To address these challenges, we propose a novel end-to-end collaborative perception framework, SCORPION, that harnesses the AI co-design of the application layer and system layer to tackle the aforementioned imperfections. SCORPION consists of three main components: lost bird’s eye view feature reconstruction (L-BEV-R) recovers lost spatial features during lossy V2X communication, while deformable spatial cross attention (DSCA) and temporal alignment (TA) compensate for localization and synchronization errors in feature fusion. Experimental results on both synthetic and real-world collaborative 3D object detection datasets demonstrate that SCORPION advances the state-of-the-art collaborative perception methods by 5.9 - 13.2 absolute AP on both standard and noisy scenarios.

Post-training quantization is a representative technique for compressing neural networks, making them smaller and more efficient for deployment on edge devices. However, an inaccessible user dataset often makes it difficult to ensure the quality of the quantized neural network in practice. In addition, existing approaches may use a single uniform bit-width across the network, resulting in significant accuracy degradation at extremely low bit-widths. To utilize multiple bit-width, sensitivity metric plays a key role in balancing accuracy and compression. In this paper, we propose a novel sensitivity metric that considers the effect of quantization error on task loss and interaction with other layers. Moreover, we develop labeled data generation methods that are not dependent on a specific operation of the neural network. Our experiments show that the proposed metric better represents quantization sensitivity, and generated data are more feasible to apply to mixed-precision quantization.

In this paper, we consider the problem of distributing patrol officers inside a building to maximize the probability of catching multiple intruders while minimizing the distance the patrol officers travel to reach the locations of the intruders. In our problem setting, the patrol officers are assisted by the information collected by a network of binary proximity sensors installed in the building. We claim that learning even common movement sub-patterns that originate due to the constrained physical environment helps to find likely locations of intruders where each major location is instrumented using a sensor node. We use a series of binary detection events to infer likely future trajectories in a real-world building. For a given set of detectable nodes on the inferred future trajectories, we aim to find the optimal patrol dispatch node location with high exposure to intruders' future appearance using patrol officers in limited numbers, ideally fewer than the intruders. In order to prevent possible crime and perform responsive defense against potential intruders, our algorithm also tries to reduce the travel distance from patrols current positions to their dispatched positions at the same time. We validate our proposed scheme in terms of detection accuracy by varying the number of intruders, robustness against missing events, and responsiveness compared to a practical baseline counterpart through real-world system experiments.