Robert Wijaya

Zhao Ning Zou, Yuhang Zhang, Robert Wijaya

Transformer-based object detectors (DETR) have shown significant performance across machine vision tasks, ultimately in object detection. This detector is based on a self-attention mechanism along with the transformer encoder-decoder architecture to capture the global context in the image. The critical issue to be addressed is how this model architecture can handle different image nuisances, such as occlusion and adversarial perturbations. We studied this issue by measuring the performance of DETR with different experiments and benchmarking the network with convolutional neural network (CNN) based detectors like YOLO and Faster-RCNN. We found that DETR performs well when it comes to resistance to interference from information loss in occlusion images. Despite that, we found that the adversarial stickers put on the image require the network to produce a new unnecessary set of keys, queries, and values, which in most cases, results in a misdirection of the network. DETR also performed poorer than YOLOv5 in the image corruption benchmark. Furthermore, we found that DETR depends heavily on the main query when making a prediction, which leads to imbalanced contributions between queries since the main query receives most of the gradient flow.

Robert Wijaya, Ngoc-Bao Nguyen, Ngai-Man Cheung

Large Vision-Language Models (LVLMs) have shown promising capabilities in understanding and generating information by integrating both visual and textual data. However, current models are still prone to hallucinations, which degrade the performance and greatly harm the user experience in real-world applications. Post-training alignment, particularly preference-tuning, is intended to align model outputs and behaviors (safety, instruction-following, style), ensuring robustness and adaptability to a wide range of tasks. The use of synthetic data for alignment, particularly in multimodal settings, remains under explored. Existing approaches typically use a strong model or a ground-truth model (CLIP) to determine positive and negative image-text data points. This paper proposes SynthAlign, a pipeline to generate and collect synthetic human-preference image-text data with optimal control built specifically for post-training alignment with DPO. At the core of the framework is the utilization of reward models as a proxy of human preference. A series of evaluation and benchmarking is provided to validate the effectiveness of the proposed framework and the resulting dataset. Notably, our framework enhanced LLaVA-1.5-7B achieved substantial POPE improvements: 87.6\% accuracy and 97.8\% precision, MMHal-Bench score increased from 2.36 to 3.49, and hallucination rate decreased from 51.0\% to 25.0\% (a 50.98\% relative reduction).