This title slide features the main heading "Dense Audio-Visual Event Localization (DAVEL)". The subtitle provides the Chinese translation: "多模态事件密集本地化研究".

Dense Audio-Visual Event Localization

(DAVEL)

多模态事件密集本地化研究

This agenda slide, titled "演示大纲" (Presentation Outline), lists five key sections of the talk. They cover DAVEL challenges versus AVEL, audio-video event localization objectives, the CCNet model with CMCC and MTGC modules, experimental results on the UnAV-100 dataset, and conclusions with future directions.

演示大纲

1. 1. 引言与挑战

阐明DAVEL挑战及与AVEL区别

2. 2. 任务目标

音频视频事件同时本地化目标

3. 3. CCNet模型

CMCC与MTGC协作模块

4. 4. 实验结果

UnAV-100数据集比较分析

5. 5. 结论与展望

贡献总结及未来方向

This section header slide introduces "Dense Audio-Visual Event Localization (DAVEL)" under Section 1: Introduction. It outlines DAVEL's key challenges—dense events and modality inconsistency—while distinguishing it from AVEL's single-event focus.

Dense Audio-Visual Event Localization (DAVEL)

1

1. 引言

DAVEL挑战：密集事件、模态不一致；区别于AVEL单事件

AVEL localizes single events in audio-video data, while DAVEL addresses challenges in dense events with difficult spatio-temporal alignment. DAVEL enables simultaneous multi-event localization for higher density and precision, shifting from single to dense multi-event processing.

DAVEL vs. AVEL

• • AVEL：单一事件音频视频本地化
• • DAVEL挑战：事件密集、时空对齐难
• • DAVEL：多事件同时定位，提高密度与精度
• • 核心区别：从单事件到密集多事件

Source: Dense Audio-Visual Event Localization (DAVEL) 引言

The slide's core focus is simultaneously localizing audio and video events. Its goal is precise spatio-temporal alignment of multimodal dense events, with outputs of event boundaries, categories, and location predictions.

2. 任务目标

• 核心：音频与视频事件同时本地化
• 目标：时空精确对齐多模态密集事件
• 输出：事件边界、类别与位置预测

This slide serves as the section header for Section 3: CCNet Model. It introduces the Cross-Modal Consistency Collaboration (CMCC) and Multi-Temporal Granularity Collaboration (MTGC) modules to achieve cross-modal consistency and multi-temporal granularity collaboration.

3. CCNet模型

3

CCNet模型

提出Cross-Modal Consistency Collaboration (CMCC)与Multi-Temporal Granularity Collaboration (MTGC)模块，实现跨模态一致性与多时粒度协作。

The CCNet architecture workflow processes input audio and video data through the CMCC module for cross-modal consistency between audio-video modalities and the MTGC module for multi-scale temporal fusion. It outputs dense event localization results, including event boundary sequences and category labels.

CCNet架构

{ "headers": [ "流程步骤", "模块功能" ], "rows": [ [ "输入音频视频", "接收原始音频和视频数据作为输入" ], [ "CMCC模块", "跨模态一致协作（Cross-Modal Consistency Collaboration），确保音频-视频模态间的一致性" ], [ "MTGC模块", "多时粒度协作（Multi-Temporal Granularity Collaboration），实现多尺度时序融合" ], [ "输出密集事件本地化", "生成密集事件边界序列和类别标签" ] ] }

Source: 流程：输入音频视频→CMCC模块（跨模态一致）→MTGC模块（多时粒度融合）→输出密集事件本地化。

This table under "Experiments and Results" compares mAP and F1 scores for baseline (69.9%, 0.72), audio-only (65.1%, 0.68), video-only (67.3%, 0.70), and CCNet (ours). CCNet outperforms others with 75.2% mAP (+5.3) and 0.78 F1 (+0.06).

4. 实验与结果

{ "headers": [ "方法", "mAP (%)", "F1" ], "rows": [ [ "基线", "69.9", "0.72" ], [ "仅音频", "65.1", "0.68" ], [ "仅视频", "67.3", "0.70" ], [ "CCNet (ours)", "75.2 (+5.3)", "0.78 (+0.06)" ] ] }

Source: UnAV-100数据集

The "性能比较" slide showcases CCNet's 75.2% mAP and 72.1% F1 score on UnAV-100. It also reports a +5.3% mAP gain versus AVEL-S.

性能比较

• 75.2%: CCNet mAP

on UnAV-100

• 72.1%: CCNet F1

on UnAV-100

• +5.3%: mAP Gain

vs. AVEL-S

The slide highlights key contributions, including the first proposal of the DAVEL task and CCNet's improvement in dense localization accuracy. It outlines future directions like more datasets and real-time applications, ending with thanks for watching.

5. 结论

**贡献

• 首提 DAVEL 任务
• CCNet 提升密集本地化精度

未来

• 更多数据集
• 实时应用

感谢观看！**

谢谢！