The detection of suicide risk in social media is a critical task with potential life-saving implications. This paper presents a study on leveraging state-of-the-art natural language processing solutions for identifying suicide risk in social media posts as a submission for the "IEEE BigData 2024 Cup: Detection of Suicide Risk on Social Media" conducted by the kubapok team. We experimented with the following configurations of transformer-based models: fine-tuned DeBERTa, GPT-4o with CoT and few-shot prompting, and fine-tuned GPT-4o. The task setup was to classify social media posts into four categories: indicator, ideation, behavior, and attempt. Our findings demonstrate that the fine-tuned GPT-4o model outperforms two other configurations, achieving high accuracy in identifying suicide risk. Notably, our model achieved second place in the competition. By demonstrating that straightforward, general-purpose models can achieve state-of-the-art results, we propose that these models, combined with minimal tuning, may have the potential to be effective solutions for automated suicide risk detection on social media.

Road maintenance inspection may be performed with the use of low-budget smartphones mounted inside a car, instead of expensive specialized vehicles with dedicated equipment. This approach, though, requires high-quality computer vision systems for processing the images. This paper describes a method developed during the Crowdsensing-based Road Damage Detection Challenge (CRDDC2022). The method focuses on efficient GPU usage during model training and creating a diversified ensemble with different architectures and data augmentation settings. My approach achieves a good result, with a 0.60 F1-score for all images and an average 0.53 F1-score across all leaderboards, which is the competition’s final metric.

A huge amount of data of various types is stored on servers and transferred over the Internet. This may pose a threat to the privacy of individuals. Encryption algorithms are developed to protect privacy. However, not all encryption mechanisms are effective. The easiest way of proving such ineffectiveness is by presenting a method of breaking the encryption scheme. In this paper, one method for matching original and encrypted images is proposed. The method was developed under the Privacy-preserving Matching of Encrypted Images shared task in the IEEE BigData 2022 Cup, and achieved second place with an accuracy of 0.6915, trailing the winning solution by only 0.0031. The proposed solution is based on Gradient Boosted Trees as implemented in CatBoost, and feature extraction of pixel value aggregates regardless of their positions. This is done using Arnold’s cat map obfuscation scheme in the encryption algorithm. Arnold’s cat map shuffles pixel positions, but leaves the image histogram unchanged. The method described in this paper almost solves two of the subtasks in the competition, reaching 0.98 accuracy for both of them; however, it does not propose a solution to the third subtask.

The popularity of computer vision systems supporting vehicle drivers and autonomous devices is increasing. Such systems require a huge quantity of annotated images. Creating such datasets is very expensive. However, it is possible to utilize synthetic images as a training dataset. In this paper, an object detection model for Vehicle Class and Orientation Detection is presented. The task set is to use solely synthetic images to train a model, which will then be evaluated on real-world images. The method takes advantage of high dataset augmentation and manual correction of training parameters regarding training statistics. The model performs surprisingly well in the detection of objects, but the assignment of classes to objects is not satisfactory. An error analysis is carried out, and propositions for future work are discussed. The described solution attained a 0.397 weighted mAP score on real-world images, achieving third place in the IEEE BigData 2022 Vehicle Class and Orientation Detection Challenge 2022.