Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Frustratingly Simple Approach for End-to-End Image Captioning

Ziyang Luo, Yadong Xi, Rongsheng Zhang, Jing Ma

TL;DR

This work introduces VC-GPT, an end-to-end image captioning framework that connects a pre-trained CLIP-ViT visual encoder with a GPT-2 language decoder through a self-ensemble cross-modal fusion mechanism. By avoiding separate object detectors and adding dedicated fusion layers, VC-GPT achieves state-of-the-art or near-state-of-the-art performance on MSCOCO, Flickr30k, and NoCaps across standard metrics, while maintaining robust generalization. The paper also provides extensive ablations, showing the importance of the self-ensemble design, the benefits of freezing the visual encoder during pretraining, and the impact of learning-rate and image-resolution choices on final caption quality. Overall, the approach offers a practical, detector-free pathway to strong cross-modal generation with broad applicability in vision-language tasks.

Abstract

Image Captioning is a fundamental task to join vision and language, concerning about cross-modal understanding and text generation. Recent years witness the emerging attention on image captioning. Most of existing works follow a traditional two-stage training paradigm. Before training the captioning models, an extra object detector is utilized to recognize the objects in the image at first. However, they require sizeable datasets with fine-grained object annotation for training the object detector, which is a daunting task. In addition, the errors of the object detectors are easy to propagate to the following captioning models, degenerating models' performance. To alleviate such defects, we propose a frustratingly simple but highly effective end-to-end image captioning framework, Visual Conditioned GPT (VC-GPT), by connecting the pre-trained visual encoder (CLIP-ViT) and language decoder (GPT2). Different from the vanilla connection method that directly inserts the cross-attention modules into GPT2, we come up with a self-ensemble cross-modal fusion mechanism that comprehensively considers both the single- and cross-modal knowledge. As a result, we do not need extra object detectors for model training. Experimental results conducted on three popular image captioning benchmarks (MSCOCO, Flickr30k and NoCaps) demonstrate that our VC-GPT achieves either the best or the second-best performance across all evaluation metrics over extensive baseline systems.

A Frustratingly Simple Approach for End-to-End Image Captioning

TL;DR

This work introduces VC-GPT, an end-to-end image captioning framework that connects a pre-trained CLIP-ViT visual encoder with a GPT-2 language decoder through a self-ensemble cross-modal fusion mechanism. By avoiding separate object detectors and adding dedicated fusion layers, VC-GPT achieves state-of-the-art or near-state-of-the-art performance on MSCOCO, Flickr30k, and NoCaps across standard metrics, while maintaining robust generalization. The paper also provides extensive ablations, showing the importance of the self-ensemble design, the benefits of freezing the visual encoder during pretraining, and the impact of learning-rate and image-resolution choices on final caption quality. Overall, the approach offers a practical, detector-free pathway to strong cross-modal generation with broad applicability in vision-language tasks.

Abstract

Image Captioning is a fundamental task to join vision and language, concerning about cross-modal understanding and text generation. Recent years witness the emerging attention on image captioning. Most of existing works follow a traditional two-stage training paradigm. Before training the captioning models, an extra object detector is utilized to recognize the objects in the image at first. However, they require sizeable datasets with fine-grained object annotation for training the object detector, which is a daunting task. In addition, the errors of the object detectors are easy to propagate to the following captioning models, degenerating models' performance. To alleviate such defects, we propose a frustratingly simple but highly effective end-to-end image captioning framework, Visual Conditioned GPT (VC-GPT), by connecting the pre-trained visual encoder (CLIP-ViT) and language decoder (GPT2). Different from the vanilla connection method that directly inserts the cross-attention modules into GPT2, we come up with a self-ensemble cross-modal fusion mechanism that comprehensively considers both the single- and cross-modal knowledge. As a result, we do not need extra object detectors for model training. Experimental results conducted on three popular image captioning benchmarks (MSCOCO, Flickr30k and NoCaps) demonstrate that our VC-GPT achieves either the best or the second-best performance across all evaluation metrics over extensive baseline systems.
Paper Structure (18 sections, 2 equations, 6 figures, 6 tables)

This paper contains 18 sections, 2 equations, 6 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Basic Single-Modal Module
  5. Cross-Modal Fusion Module
  6. Training Objective
  7. Experiment Setup
  8. Experiment Setup
  9. Experiment Results
  10. Comparing with Vanilla Framework
  11. Comparing with End-to-End Baseline Systems
  12. Comparing with Two-Stage Training Baseline Systems
  13. Qualitative Evaluation
  14. Cross-Attention Map Visualization
  15. Extra Empirical Exploration
  16. ...and 3 more sections

Figures (6)

  • Figure 1: Comparing the end-to-end image captioning paradigm with the previous two-stage training paradigm.
  • Figure 2: Comparing the differences between the Vanilla Framework and our VC-GPT Framework.
  • Figure 3: Self-Ensemble mechanism can comprehensively consider both single- and cross-modal knowledge, leading to better performance. All logits are generated by our VC-GPT.
  • Figure 4: Image captioning examples of our VC-GPT and OSCAR for the first 4 images in the MSCOCO test set. Blue indicates that such caption or word is inaccurate, after comparing with the ground truth. For simplicity, we only show the first two golden captions per image.
  • Figure 5: Visualization of the cross-attention maps in the final fusion layer of text tokens. If the color of a attention block is lighter, the cross-attention score is higher.
  • ...and 1 more figures