Fine-grained Image Captioning with CLIP Reward

Abstract

Modern image captioning models are usually trained with text similarity objectives. However, since reference captions in public datasets often describe the most salient common objects, models trained with the text similarity objectives tend to ignore specific and detailed aspects of an image that distinguish it from others. Towards more descriptive and distinctive caption generation, we propose to use CLIP, a multimodal encoder trained on huge image-text pairs from the web, to calculate multi-modal similarity and use it as a reward function. We also propose a simple finetuning strategy of CLIP text encoder to improve grammar that does not require extra text annotation. This completely eliminates the need for reference captions during the reward computation. To comprehensively evaluate descriptive captions, we introduce FineCapEval, a new dataset for caption evaluation with fine-grained criteria: overall, background, object, relations. In our experiments on text-to-image retrieval and FineCapEval, the proposed CLIP-guided model generates more distinctive captions than the CIDEroptimized model. We also show that our unsupervised grammar finetuning of the CLIP text encoder alleviates the degeneration problem of the naive CLIP reward. Lastly, we show human analysis where the annotators strongly prefer CLIP reward to CIDEr and MLE objectives on diverse criteria. 1

Blind Hyperspectral Image Denoising with Degradation Information Learning

“…Unsupervised metrics directly capture the similarity between input images and generated sentences. A typical metric, CLIP-S [53], quantifies the cosine similarity between features of images and sentence extracted from pretrained CLIP model.…”

“…Recent research [53] demonstrates that matching scores the image and text based on CLIP can increase the diversity and accuracy of captions generated. This enhancement is achieved through direct measurement of the correlation between the input image and the generated sentences.…”

“…While calculating the loss function L, it is important to note that some operations involved in the process are non-differentiable, such as sampling the probability distribution to retrieve words. Thus, to ensure the update of the gradient, following [54,53], we optimize the model with REINFORCE algorithm [55] with a self-critical baseline.…”

Cross-Domain Image Captioning with Discriminative Finetuning

“…Among the approaches more closely related to ours, the system of Yu et al [51] uses a ClipCap-like system for caption generation, and CLIP to measure image-caption similarity, focusing on generating captions in multiple styles. Cho et al [8] use CLIP to finetune a pre-trained captioner. Like [51], they use the CLIP-Score image-caption similarity measure [16] as a reward signal, rather than a discriminative objective like the one we adopt.…”

DeAR: Debiasing Vision-Language Models with Additive Residuals

“…Deep learning-based vision-language models (VLMs) [45] unify text and visual data into a common representation and reduce the computing cost of training for specific computer vision [19] and visually-grounded linguistic tasks [12,47]. VLMs are trained with a large amount of data with the aim of matching image and text representations for image-caption pairs to capture diverse visual and linguistic concepts.…”