Quantization-Aware Training (QAT) for vision transformers relies on expensive retraining to recover accuracy loss in non-linear layer quantization, which limits their use in resource-constrained environments. In contrast, existing Post-Training Quantization (PTQ) methods either partially quantize non-linear functions or adjust activation distributions to maintain accuracy but fail to achieve fully integer-only inference. In this paper, we introduce IPTQ-ViT, a PTQ framework for fully integer-only vision transformers without retraining. We present approximation functions: a polynomial-based GELU optimized for vision data and a bit-shifting-based Softmax designed to improve approximation accuracy in PTQ. In addition, we propose a unified metric integrating quantization sensitivity, perturbation, and computational complexity to select the optimal approximation function per activation layer. IPTQ-ViT consistently outperforms previous PTQ methods, achieving up to 6.44%p (avg. 1.78%p) top-1 accuracy improvement for image classification and 1.0 mAP for object detection under W8A8 and W4A8. IPTQ-ViT achieves accuracy and latency comparable to integer-only QAT methods.

Unmanned aerial vehicles (UAVs) are being used for monitoring natural disasters. To promptly assess the severity of the disaster, it is advantageous to analyze the disaster scenes with the on-board computer in real time. Various AI segmentation models with high accuracy are developed mainly for offline processing. They require significant memory capacity and computational power. However, on-device AI has the challenge of compressing high-precision models due to the limited memory size and the small computation power. In this research, we develop a lightweight disaster semantic segmentation model for UAV on-device intelligence. From a simple FANet as our baseline, we apply various optimization, such as compact backbone, lightweight attention block, quantization, knowledge distillation, and post processing. With our optimized model, we can reduce 84.2% of the inference time while achieving 0.5% increase in accuracy compared to the baseline model.

Smartwatches have gained significant popularity in recent years. One major use of smartwatches is notification checking as an extended display. Smartwatch use provides an opportunity for energy saving because it affords a reduction in the frequency of smartphone use. However, sometimes user input is limited in smartwatches due to its small screen size and users are forced to use their smartphones to fully read notifications. In such cases, the advantage of an extended display in the smartwatch diminishes because users need to access their smartphones, which causes extra smartphone battery consumption. We define phone-preferable notification that requires a user to take further actions, such as checking detailed content and replying to a message. Given that phone-preferable notifications are likely to be handled on smartphones, it is possible to defer notification delivery to smartwatches. In this paper, we develop a novel notification manager, called PASS that automatically defers phone-preferable notifications and piggybacks them on watch-preferable notifications. For model building and evaluation, we collect 15,659 notifications in-the-wild from 11 users for approximately 31 days. In addition, for approximately two months we gather self-reported data from five users regarding which devices were used to respond to notifications. The results show that PASS can save daily battery for smartwatches and daily battery for smartphones up to 43.5 and 0.9 percent without introducing any noticeable negative results on user experiences.

Image-based convolutional neural network(CNN) algorithms are spreading across a variety of applications. In particular, an autonomous vehicle recognizes objects and the surrounding situation using the CNN models. CNN models for image classification are trained using clear image dataset, so they are not robust to grayscale images or noise-intensive data. Therefore, there is a risk of an accident because the quality of the input image drops rapidly during night driving. The region that is revealed by the headlight can have colors, but in the shaded area it has a brightness that is not enough to get color values. We intend to increase the safety of autonomous driving by coloring this region of interest(ROI).

In order to guarantee task schedulability in real-time embedded systems, analysis on cache and cache behavior at task preemption is required because cache misses cause unpredictable delays in task execution time. Nevertheless, studies on cache behavior and its effects in preemptive multitasking systems have been conducted mainly with simulation and analytic methods, which may induce pessimistic scheduling policy. In this paper, we propose a measurement framework to diagnose the cache-related preemption delay (CRPD) on LEON3 SPARC V8 processor to understand its effect with execution time analysis. The experimental results from several benchmark program execution on the framework show that the time delay and penalty for cache operations are not constant. We were able to identify that each software has a unique degree of cache susceptibility with regard to CRPD. We conclude that this experiment data can be used to prevent real-time tasks from missing the deadline, which is not tolerable in real-time constraints of spacecraft computers, by applying it on real-time scheduling. The results may be useful to the susceptibility analysis of real-time embedded software. The measured CRPD on a specified real target processor can be used to assess the schedulability of task sets to bound WCET on task preemption condition.

Mission critical embedded software for autonomous operation requires high development cost due to its long development cycle. One of the potential solutions for reducing the cost is to reuse the software developed at previous missions. Virtual machine platform such as JVM is a good example to provide code portability across various missions. Flight software in aerospace field is adopting this concept to improve reusability and eventually to reduce development cost. In this paper, we propose a Lua-based virtualization environment for spacecraft flight software. Flight software for spacecraft control consists of a few tasks that are highly autonomous. Lua is chosen as the script language for programming the control tasks. Though Lua was designed with simplicity and portability, it only supports multithreading with collaborative coroutines. To support preemptive multitasking, we implement time slicing coroutines as spacecraft control processes. New coroutine scheduler is devised and time slicing functionality is added into the scheduler. Scheduler locking and message passing with external flight software are also implemented. Instead of modifying the Lua interpreter, we have exploited the debug support APIs for our implementation. For evaluation, we have implemented the flight software virtualization environment on the flight computer. Accuracy of the time slicing scheduler is also analyzed.

Development of the space industry has led to a diversity of specialized electronics-units implemented on space system. However, this provoked new problems such as increased system's complexity and the size. As a solution to the problem, the Integrated Modular Avionics (IMA) architecture and ARINC 653 standard have been suggested. This paper introduces a Linux-based memory efficient ARINC 653 partition scheduler, which employs fixed-priority preemptive scheduling algorithm for partitions scheduling. The implemented partition scheduler has small memory footprint for each partition and produces low partition switching overhead. The prototype was executed on a LEON4 processor, which is the European Space Agency (ESA) Next Generation Multicore Processor (NGMP) in the space sector. In evaluation, we analyzed execution trace, memory footprint and scheduling overhead.

In spite of the incredible market penetration of smartphones and rapid growth of App markets, their utility has been restricted by their battery capacity. Smartphone users are frustrated by applications consuming too much power. To avoid this experience, they may have to profile energy consumption on the phone. However, it is difficult for a user to analyze energy consumption of his phone because the process requires deep understanding of the interaction between the hardware and software. Therefore, market curators need to examine applications for providing power consumption data before accepting them into App market. Unfortunately, current App markets mainly employ basic functional tests and checking policy on the submitted applications. Moreover, considering the exponentially increasing daily amount of new Apps posted on the App market, it is not feasible to perform the process manually. Therefore, if we can provide an automated energy estimation tool for App market curators, users can have energy consumption data in addition to the current functional description. In this poster, we propose a framework for automated power estimation of android apps.

The paper introduces a robust image watermark based on an invariant image feature vector. Normalized Zernike moments of an image are used as the vector. The watermark is generated by modifying the vector. The watermark signal is designed with Zernike moments. The signal is added to the cover image in the spatial domain after the reconstruction process. We extract the feature vector from the modified image and use it as the watermark. The watermark is detected by comparing the computed Zernike moments of the test image and the given watermark vector. Rotation invariance is achieved by taking the magnitude of the Zernike moments. An image normalization method is used for scale and translation invariance. The robustness of the proposed method is demonstrated and tested using Stirmark 3.1. The test results show that our watermark is robust with respect to geometrical distortions and compression.