Felix Becker

Tuna Tuncer, Felix Becker, Thomas Pfeil

Chunk-wise autoregressive video diffusion models rely on a KV cache of previously generated chunks to avoid redundant computation, but this cache quickly becomes a memory bottleneck as videos grow longer. Methods that quantize the KV cache to low bitwidths reduce memory pressure but degrade video quality. We show that a key driver of this degradation is a systematic bias in attention weights: due to the convexity of the exponential in softmax attention, quantization noise inflates the contribution of cached keys, a phenomenon we call the Jensen bias. This effect causes quantized keys to steal attention mass from the unquantized current chunk. We derive a per-attention-score correction that removes this bias in expectation, computed on the fly from the quantization step sizes of the cached keys and the query norm. Using a second-order Taylor approximation, the additional computational overhead is negligible, and no additional memory is needed alongside the cache. Evaluated on MAGI-1, SkyReels-V2, and HY-WorldPlay at INT2 quantization, our correction recovers most of the quality lost to aggressive quantization, reaching near-BF16 video quality, and can outperform INT4 quantization while using 50% less memory.

Dominik Notz, Felix Becker, Thomas Kühbeck et al.

Collecting realistic driving trajectories is crucial for training machine learning models that imitate human driving behavior. Most of today's autonomous driving datasets contain only a few trajectories per location and are recorded with test vehicles that are cautiously driven by trained drivers. In particular in interactive scenarios such as highway merges, the test driver's behavior significantly influences other vehicles. This influence prevents recording the whole traffic space of human driving behavior. In this work, we present a novel methodology to extract trajectories of traffic objects using infrastructure sensors. Infrastructure sensors allow us to record a lot of data for one location and take the test drivers out of the loop. We develop both a hardware setup consisting of a camera and a traffic surveillance radar and a trajectory extraction algorithm. Our vision pipeline accurately detects objects, fuses camera and radar detections and tracks them over time. We improve a state-of-the-art object tracker by combining the tracking in image coordinates with a Kalman filter in road coordinates. We show that our sensor fusion approach successfully combines the advantages of camera and radar detections and outperforms either single sensor. Finally, we also evaluate the accuracy of our trajectory extraction pipeline. For that, we equip our test vehicle with a differential GPS sensor and use it to collect ground truth trajectories. With this data we compute the measurement errors. While we use the mean error to de-bias the trajectories, the error standard deviation is in the magnitude of the ground truth data inaccuracy. Hence, the extracted trajectories are not only naturalistic but also highly accurate and prove the potential of using infrastructure sensors to extract real-world trajectories.