Kenan E. Ak

Taehoon Kim, Pyunghwan Ahn, Sangyun Kim et al. · nvidia, utoronto

In this report, we introduce NICE (New frontiers for zero-shot Image Captioning Evaluation) project and share the results and outcomes of 2023 challenge. This project is designed to challenge the computer vision community to develop robust image captioning models that advance the state-of-the-art both in terms of accuracy and fairness. Through the challenge, the image captioning models were tested using a new evaluation dataset that includes a large variety of visual concepts from many domains. There was no specific training data provided for the challenge, and therefore the challenge entries were required to adapt to new types of image descriptions that had not been seen during training. This report includes information on the newly proposed NICE dataset, evaluation methods, challenge results, and technical details of top-ranking entries. We expect that the outcomes of the challenge will contribute to the improvement of AI models on various vision-language tasks.

Gwang Gook Lee, Kenan Emir Ak, Jay Mohta et al.

Vision Language Models (VLMs) are increasingly used in place of traditional OCR pipelines for document understanding. In this paper, we show they do not always act as faithful transcribers: when text is imperfect, they often tend to rewrite it into a more plausible form - a behavior that clean-text OCR benchmarks cannot detect. We introduce FaithC4, a multilingual perturbation benchmark of 1,455 single-page documents (English, Chinese, Korean) with three perturbation families: scramble, random substitution, and visually similar substitution. We use the benchmark to evaluate 15 systems spanning general-purpose VLMs, OCR-specialized VLMs, and traditional OCR pipelines. These three categories differ in WER degradation under perturbation: general-purpose VLMs degrade by up to 4.5 points, OCR-specialized VLMs by 0.2-2 points, and traditional OCR by less than 0.6 points on English. Probing Qwen3-VL-4B layer-by-layer, we identify a consistent pattern: rewriting fires only when a perturbed word's final layer FFN representation stays close to the original encoding; when the representation diverges sufficiently, the model transcribes faithfully. Word length affects rewriting rate: short words (4-6 characters) are rewritten up to 10% of the time, with a sharp cutoff at 8 characters above which rewriting drops to 0%.

Kenan E. Ak, Joo Hwee Lim, Ying Sun et al.

The focus of this paper is on the problem of image retrieval with attribute manipulation. Our proposed work is able to manipulate the desired attributes of the query image while maintaining its other attributes. For example, the collar attribute of the query image can be changed from round to v-neck to retrieve similar images from a large dataset. A key challenge in e-commerce is that images have multiple attributes where users would like to manipulate and it is important to estimate discriminative feature representations for each of these attributes. The proposed FashionSearchNet-v2 architecture is able to learn attribute specific representations by leveraging on its weakly-supervised localization module, which ignores the unrelated features of attributes in the feature space, thus improving the similarity learning. The network is jointly trained with the combination of attribute classification and triplet ranking loss to estimate local representations. These local representations are then merged into a single global representation based on the instructed attribute manipulation where desired images can be retrieved with a distance metric. The proposed method also provides explainability for its retrieval process to help provide additional information on the attention of the network. Experiments performed on several datasets that are rich in terms of the number of attributes show that FashionSearchNet-v2 outperforms the other state-of-the-art attribute manipulation techniques. Different than our earlier work (FashionSearchNet), we propose several improvements in the learning procedure and show that the proposed FashionSearchNet-v2 can be generalized to different domains other than fashion.

Kenan E. Ak, Ning Xu, Zhe Lin et al.

Autoregressive models recently achieved comparable results versus state-of-the-art Generative Adversarial Networks (GANs) with the help of Vector Quantized Variational AutoEncoders (VQ-VAE). However, autoregressive models have several limitations such as exposure bias and their training objective does not guarantee visual fidelity. To address these limitations, we propose to use Reinforced Adversarial Learning (RAL) based on policy gradient optimization for autoregressive models. By applying RAL, we enable a similar process for training and testing to address the exposure bias issue. In addition, visual fidelity has been further optimized with adversarial loss inspired by their strong counterparts: GANs. Due to the slow sampling speed of autoregressive models, we propose to use partial generation for faster training. RAL also empowers the collaboration between different modules of the VQ-VAE framework. To our best knowledge, the proposed method is first to enable adversarial learning in autoregressive models for image generation. Experiments on synthetic and real-world datasets show improvements over the MLE trained models. The proposed method improves both negative log-likelihood (NLL) and Fréchet Inception Distance (FID), which indicates improvements in terms of visual quality and diversity. The proposed method achieves state-of-the-art results on Celeba for 64 $\times$ 64 image resolution, showing promise for large scale image generation.