5%>100%: Breaking Performance Shackles of Full Fine-Tuning on Visual Recognition Tasks

TL;DR

The paper questions the continued supremacy of full fine-tuning for visual transfer learning and introduces Mona-tuning, a convolution-based adapter framework with multi-cognitive visual filters and input distribution control. Mona inserts vision-focused adapters after fixed backbone layers (Swin blocks) and uses input normalization to align feature distributions, enabling delta-tuning that can outperform full fine-tuning on both simple and dense prediction tasks. Across COCO, ADE20K, Pascal VOC, DOTA/STAR, Flowers102, Oxford Pets, and VOC2007, Mona achieves consistent gains, including approximately a 1% AP improvement on COCO, while drastically reducing the number of trainable parameters. This work suggests that, with carefully designed visual adapters, full fine-tuning is no longer universally optimal for transferring large vision models to diverse tasks, offering practical benefits in efficiency and performance.

Abstract

Pre-training & fine-tuning can enhance the transferring efficiency and performance in visual tasks. Recent delta-tuning methods provide more options for visual classification tasks. Despite their success, existing visual delta-tuning art fails to exceed the upper limit of full fine-tuning on challenging tasks like object detection and segmentation. To find a competitive alternative to full fine-tuning, we propose the Multi-cognitive Visual Adapter (Mona) tuning, a novel adapter-based tuning method. First, we introduce multiple vision-friendly filters into the adapter to enhance its ability to process visual signals, while previous methods mainly rely on language-friendly linear filters. Second, we add the scaled normalization layer in the adapter to regulate the distribution of input features for visual filters. To fully demonstrate the practicality and generality of Mona, we conduct experiments on multiple representative visual tasks, including instance segmentation on COCO, semantic segmentation on ADE20K, object detection on Pascal VOC, oriented object detection on DOTA/STAR, and image classification on three common datasets. Exciting results illustrate that Mona surpasses full fine-tuning on all these tasks, and is the only delta-tuning method outperforming full fine-tuning on the above various tasks. For example, Mona achieves 1% performance gain on the COCO dataset compared to full fine-tuning. Comprehensive results suggest that Mona-tuning is more suitable for retaining and utilizing the capabilities of pre-trained models than full fine-tuning. The code will be released at https://github.com/Leiyi-Hu/mona.

Figures (3)

• Figure 1: Comparisons of our method with full fine-tuning and recent delta-tuning art on representative visual tasks. Blue dashed line is the performance of full fine-tuning on ADE20K and COCO. The proposed Mona outperforms full fine-tuning on representative visual tasks, which promotes the upper limit of previous delta-tuning art. The results demonstrate that the adapter-tuning paradigm can replace full fine-tuning and achieve better performance in common visual tasks. Full fine-tuning may no longer be the only preferred solution for transfer learning in the future.
• Figure 2: Left: The proposed Mona-tuning. We add Mona after MSA and MLP in each SwinBlock. The proposed method fixes the parameters of pre-trained layers and updates the parameters of Mona. Right: Details of Mona. Mona has a scaled LayerNorm before the down projection. A multi-cognitive convolutional filter group and an aggregation filter are behind the down projection. We add skip-connections at four places inside Mona to strengthen its adaptation capabilities. Mona enables the adapter-based fine-tuning paradigm to outperform full fine-tuning in typical visual tasks comprehensively.
• Figure 3: Loss curves. Among all the methods, the proposed method converges faster and significantly exceeds the full fine-tuning.