Temporal Preservation over Processing: Diagnosing and Designing Spatiotemporal Single-Stage Video Detectors
Authors: Karam Tomotaki-Dawoud, Anna Hilsmann, Peter Eisert, Sebastian Bosse
Organizations: Fraunhofer HHI, Berlin, Germany
Abstract
Single-stage video object detectors are increasingly deployed in time-critical applications, yet it remains unclear whether these models genuinely reason over temporal context or merely exploit a single informative frame-a gap hidden by standard metrics, which reward correct predictions regardless of how they are reached. We address this from two complementary directions: first, we propose TemporalLens, a model-agnostic diagnostic framework probing temporal dependence through controlled perturbations, structured occlusions, temporal shuffling, redundancy injection, and resolution degradation, revealing whether a detector actually uses information across time. Applied to stacked-frame 2D detectors and our YOLO-3D architecture, it exposes behavioural differences invisible to mAP: stacked 2D models collapse when the target frame is removed, while spatiotemporal models recover predictions from earlier frames, a signature of real temporal reliance. Second, we detail YOLO-3D, a modular real-time spatiotemporal detector built on YOLOv8, and show that simply preserving temporal depth through the backbone is the dominant performance driver (+3.7 pp mAP@50 at 32 frames averaged across scales). Together, the diagnostics and architecture turn "does this detector reason over time?" into a measurable, actionable question.
Modern object detectors are static, fixed-depth networks optimized for a single operating point, requiring separate models for different deployment scenarios. We present an any-depth detection framework that enables a single network to span a continuous range of accuracy--efficiency trade-offs by controlling depth at inference time without retraining. Each backbone and neck stage is divided into an essential path, which always executes, and a skippable refinement path; this decomposition preserves the full multi-scale feature hierarchy at every depth configuration, unlike conventional early exiting that discards entire stages. To train such a network, jointly optimizing many sub-networks of varying depth introduces conflicting gradient signals. We address this via self-distillation between only the two extremes, with prediction-level and feature-level alignment losses that enforce stage-wise modularity, ensuring the outputs of each stage remain compatible regardless of the paths taken. Instantiated on RT-DETR and YOLOv12, our full-depth configurations match or surpass their respective SOTA baselines with negligible parameter overhead, while the most efficient configurations achieve up to 1.82× speedup at a cost of only 2.0 AP, all from a single set of weights.
Spatio-temporal action detection in videos requires jointly localizing actors in space and identifying action boundaries over time. A common challenge is constructing temporally stable action tubes, as frame-level detectors often suffer from jitter, fragmentation, and imprecise temporal localization. Many recent approaches address this by introducing heavy spatio-temporal transformers or optical-flow-based pipelines, leading to high computational cost and limited scalability. We propose TubeLite, a lightweight framework for spatio-temporal action detection that focuses on stable tube construction and boundary-aware temporal modeling. TubeLite represents each actor as a tube, defined as a sequence of bounding boxes associated with a single actor over time, and explicitly enforces temporal consistency at both the spatial and semantic levels. The method combines low-jitter actor detection, Gaussian-weighted actor feature extraction, efficient short-term temporal propagation, and a boundary-focused temporal prediction head, while avoiding optical flow and large-scale temporal attention. Despite its compact design, TubeLite achieves strong video-level localization performance. It improves Video-mAP@0.5 by 4.5 and 7.1 percentage points over the best compared method on the MultiSports and UCF101-24 datasets, respectively, with substantially fewer parameters and floating-point operations than transformer-based alternatives, demonstrating that effective spatio-temporal action detection can be obtained through principled, lightweight temporal modeling.
Ali Soltaninezhad, Melissa Cote, Alejandro Rico Espinosa +2
Fine-grained spatio-temporal understanding is essential for video reasoning and embodied AI. Yet, while Multimodal Large Language Models (MLLMs) master static semantics, their grasp of temporal dynamics remains brittle. We present TimeBlind, a diagnostic benchmark for compositional spatio-temporal understanding. Inspired by cognitive science, TimeBlind categorizes fine-grained temporal understanding into three levels: recognizing atomic events, characterizing event properties, and reasoning about event interdependencies. Unlike benchmarks that conflate recognition with temporal reasoning, TimeBlind leverages a minimal-pairs paradigm: video pairs share identical static visual content but differ solely in temporal structure, utilizing complementary questions to neutralize language priors. Evaluating over 20 state-of-the-art MLLMs (e.g., GPT-5, Gemini 3 Pro) on 600 curated instances (2400 video-question pairs), reveals that the Instance Accuracy (correctly distinguishing both videos in a pair) of the best performing MLLM is only 48.2%, far below the human performance (98.2%). These results demonstrate that even frontier models rely heavily on static visual shortcuts rather than genuine temporal logic, positioning TimeBlind as a vital diagnostic tool for next-generation video understanding. Dataset and code are available at https://baiqi-li.github.io/timeblind_project/ .