AVESFormer: Efficient Transformer Design for Real-Time Audio-Visual Segmentation

TL;DR

This work tackles the real-time constraint in audio-visual segmentation by identifying attention dissipation in cross-attention Fusion and proposing an efficient decoder design. It introduces AVESFormer, which combines a Prompt Query Generator (PQG) to enrich audio-conditioned queries and an EarLy Focus (ELF) decoder that replaces early transformer attention with convolution to reduce cost while preserving performance. The approach achieves a leading trade-off between accuracy and inference speed on AVSBench with S4, MS3, and AVSS, reporting mIoU of 79.9% on S4, 57.9% on MS3, and 31.2% on AVSS, with substantial parameter and speed advantages. Ablation studies demonstrate the effectiveness of PQG and ELF, while noting limitations such as the heavy audio backbone and lack of temporal modeling for real-time scenarios.

Abstract

Recently, transformer-based models have demonstrated remarkable performance on audio-visual segmentation (AVS) tasks. However, their expensive computational cost makes real-time inference impractical. By characterizing attention maps of the network, we identify two key obstacles in AVS models: 1) attention dissipation, corresponding to the over-concentrated attention weights by Softmax within restricted frames, and 2) inefficient, burdensome transformer decoder, caused by narrow focus patterns in early stages. In this paper, we introduce AVESFormer, the first real-time Audio-Visual Efficient Segmentation transformer that achieves fast, efficient and light-weight simultaneously. Our model leverages an efficient prompt query generator to correct the behaviour of cross-attention. Additionally, we propose ELF decoder to bring greater efficiency by facilitating convolutions suitable for local features to reduce computational burdens. Extensive experiments demonstrate that our AVESFormer significantly enhances model performance, achieving 79.9% on S4, 57.9% on MS3 and 31.2% on AVSS, outperforming previous state-of-the-art and achieving an excellent trade-off between performance and speed. Code can be found at https://github.com/MarkXCloud/AVESFormer.git.

Figures (10)

• Figure 1: Illustration of attention dissipation. The cross-attention matrix fails to distinguish different tokens (left). One potential solution is to expand the audio feature into several tokens (right).
• Figure 2: mIoU (%) vs. Inference Latency (ms) on S4 (left), MS3 (middle) and AVSS (right) compared with other popular methods. Latency is measured on a single Nvidia RTX 3090 GPU. AVESFormer achieves the best trade-off between performance and inference speed.
• Figure 3: Runtime profiling of the AVSegFormer gao2024avsegformer. Inference time is dominated by transformer architecture.
• Figure 4: The overview of AVESFormer. Audio and visual backbones extract corresponding features. The prompt query generator addresses the attention dissipation problem by inserting the audio feature on top of a set of learnable parameters to generate audio-conditioned queries. The ELF decoder processes local features using convolution blocks in the early stages. Finally, the transformer blocks interact with high-level audio-visual features to generate fused features.
• Figure 5: Illustration of prompt query generator. The audio feature is treated as prompt and discarded in output.