docs: define speaker-grounded floor experiment plan
This commit is contained in:
114
README.md
114
README.md
@@ -1,8 +1,45 @@
|
|||||||
# Omni-only Turn-Taking Experiment
|
# Speaker-Grounded Omni Floor-State Forecasting
|
||||||
|
|
||||||
This project tests whether **Qwen2.5-Omni-3B can replace both the handcrafted
|
This project studies whether a strictly causal, end-to-end pretrained
|
||||||
audio/visual feature extractors and the downstream RNN/Transformer used by
|
**Qwen2.5-Omni-3B** backbone contains speaker-grounded audiovisual evidence,
|
||||||
EgoSpeak** for egocentric turn-taking anticipation.
|
and whether that evidence can improve the **floor-entry and floor-yield
|
||||||
|
actions of an invisible egocentric wearer** in natural multiparty
|
||||||
|
conversations.
|
||||||
|
|
||||||
|
Replacing EgoSpeak's handcrafted encoders and downstream GRU/Transformer is
|
||||||
|
retained as a Stage-0 representation baseline. It is no longer the final paper
|
||||||
|
question. The central test is behavioral: at matched wearer-onset recall,
|
||||||
|
correct person-aligned vision should reduce false wearer entries while two
|
||||||
|
other people are conversing, without reducing responses addressed to the
|
||||||
|
wearer or degrading wearer yield decisions.
|
||||||
|
|
||||||
|
## Current paper question and experiment plan (2026-08-30)
|
||||||
|
|
||||||
|
The full staged plan is in
|
||||||
|
[the Omni speaker-floor experiment plan](docs/omni_speaker_floor_experiment_plan_2026-08-30.md).
|
||||||
|
It separates four questions that must not be conflated:
|
||||||
|
|
||||||
|
1. Does a frozen global Omni state encode aggregate wearer/others activity?
|
||||||
|
2. Can Omni features identify **which visible person** is speaking, rather
|
||||||
|
than merely detect that somebody is speaking?
|
||||||
|
3. Does anonymous person-wise causal fusion reduce third-party false entries
|
||||||
|
at matched wearer onset recall?
|
||||||
|
4. Does any suppression preserve addressed-to-wearer response recall/latency
|
||||||
|
and wearer yield behavior, rather than making the model globally passive?
|
||||||
|
|
||||||
|
The immediate order is: freeze the event protocol and competing-risk
|
||||||
|
attribution; run speaker-grounding probes and wrong-person/time-shift controls;
|
||||||
|
re-evaluate `G`, `G+A`, and `G+A+pooled-V`; then train an anonymous person-wise
|
||||||
|
causal side branch. A person-memory adapter or Thinker LoRA is a second-stage
|
||||||
|
experiment only after the simpler branch passes those diagnostic gates.
|
||||||
|
|
||||||
|
The literature boundary and novelty threats are recorded in
|
||||||
|
[the multiparty egocentric floor gap audit](docs/multiparty_egocentric_floor_gap_audit_2026-08-30.md).
|
||||||
|
The precise target is the intersection of egocentric wearer action, strict
|
||||||
|
causality, speaker/addressee grounding, both entry and yield, and event-level
|
||||||
|
evaluation. This repository does **not** claim the first multiparty AV
|
||||||
|
turn-taking system, the first online multiparty EOT detector, or that a frozen
|
||||||
|
Omni backbone is task-level end-to-end training.
|
||||||
|
|
||||||
The persistent paper index and Chinese summaries are maintained in
|
The persistent paper index and Chinese summaries are maintained in
|
||||||
[literature_notes](literature_notes/README.md).
|
[literature_notes](literature_notes/README.md).
|
||||||
@@ -22,7 +59,7 @@ horizon curves are archived in
|
|||||||
It keeps causal-valid summaries separate from full-timeline curves and records
|
It keeps causal-valid summaries separate from full-timeline curves and records
|
||||||
the head-capacity/context caveats for comparisons with G+A and G+A+V.
|
the head-capacity/context caveats for comparisons with G+A and G+A+V.
|
||||||
|
|
||||||
The next research plan is
|
The earlier engineering plan is
|
||||||
[speaker-aware Thinker and wearer-transition evaluation](docs/speaker_aware_thinker_plan_2026-08-28.md)
|
[speaker-aware Thinker and wearer-transition evaluation](docs/speaker_aware_thinker_plan_2026-08-28.md)
|
||||||
(2026-08-28): person-organized video chunks, frozen-Thinker local-evidence
|
(2026-08-28): person-organized video chunks, frozen-Thinker local-evidence
|
||||||
interventions, and onset/offset plus third-party-conversation diagnostics.
|
interventions, and onset/offset plus third-party-conversation diagnostics.
|
||||||
@@ -34,7 +71,7 @@ Thinker LoRA on 8×L20, with matched-onset-recall false-entry evaluation and
|
|||||||
addressee-specific safeguards. It also records related-work and input-permission
|
addressee-specific safeguards. It also records related-work and input-permission
|
||||||
limits; none of these new variants has been implemented or measured yet.
|
limits; none of these new variants has been implemented or measured yet.
|
||||||
|
|
||||||
The main experiment is deliberately simple:
|
The Stage-0 representation experiment is deliberately simple:
|
||||||
|
|
||||||
```text
|
```text
|
||||||
causal audio + video up to time t
|
causal audio + video up to time t
|
||||||
@@ -51,9 +88,11 @@ causal audio + video up to time t
|
|||||||
future logits [10 horizons, 3 classes]
|
future logits [10 horizons, 3 classes]
|
||||||
```
|
```
|
||||||
|
|
||||||
There is **no RNN after Qwen**. The Omni Thinker is the temporal model. The
|
There is **no RNN after Qwen in this Stage-0 baseline**. The Omni Thinker is
|
||||||
only supervised component in the first experiment is a small `10 x 3`
|
its temporal model, and its only supervised component is a small `10 x 3`
|
||||||
classification head.
|
classification head. Later person-wise GRU/TCN branches and local-memory
|
||||||
|
adapters are controlled methods for testing speaker grounding; they do not
|
||||||
|
change what the Stage-0 result measures.
|
||||||
|
|
||||||
## Current status
|
## Current status
|
||||||
|
|
||||||
@@ -66,6 +105,8 @@ classification head.
|
|||||||
| `TurnPredictionHead` | Trained and evaluated | Map `[2048]` to `[10, 3]` logits |
|
| `TurnPredictionHead` | Trained and evaluated | Map `[2048]` to `[10, 3]` logits |
|
||||||
| `train_turn_head.py` | Verified on EasyCom | Train the frozen-Omni anticipation head with the official EasyCom split |
|
| `train_turn_head.py` | Verified on EasyCom | Train the frozen-Omni anticipation head with the official EasyCom split |
|
||||||
| `infer_turn_head.py` | Verified on EasyCom | Save EgoSpeak-format predictions and reproduce per-horizon AP/mAP |
|
| `infer_turn_head.py` | Verified on EasyCom | Save EgoSpeak-format predictions and reproduce per-horizon AP/mAP |
|
||||||
|
| `extract_turn_local_features.py` | Verified on EasyCom; Ego4D extraction in progress | Extract person-crop and recent-audio Qwen sidecars |
|
||||||
|
| `evaluate_floor_state_changes.py` | Implemented; attribution revision required | Diagnose wearer entry/yield events from saved predictions |
|
||||||
| `eval_omni_turn_lora.py` | Implemented, remote validation pending | Evaluate Thinker LoRA from dense AV caches with the same AP/mAP protocol |
|
| `eval_omni_turn_lora.py` | Implemented, remote validation pending | Evaluate Thinker LoRA from dense AV caches with the same AP/mAP protocol |
|
||||||
|
|
||||||
The full Thinker input path, embedded audio, direct video decoding,
|
The full Thinker input path, embedded audio, direct video decoding,
|
||||||
@@ -1687,28 +1728,40 @@ Reference Table 2 targets from the EgoSpeak paper:
|
|||||||
Do not compare these numbers with EgoSpeak Table 3 target-speaker-only AP;
|
Do not compare these numbers with EgoSpeak Table 3 target-speaker-only AP;
|
||||||
Table 3 uses a different metric.
|
Table 3 uses a different metric.
|
||||||
|
|
||||||
## Experiments that demonstrate Omni's semantic advantage
|
## Current paper experiment ladder
|
||||||
|
|
||||||
The headline comparison alone cannot establish that improvements come from
|
The headline comparison cannot establish speaker grounding or social
|
||||||
semantics. Run the following controlled experiments:
|
understanding. Experiments therefore proceed through explicit gates:
|
||||||
|
|
||||||
1. **Omni audio vs Omni audio+video**: measure whether visual grounding helps.
|
1. **P0 — protocol correctness:** preserve physical 5 Hz timing and causal
|
||||||
2. **Per-horizon curve**: semantic completion should be especially visible at
|
masks, then prevent an other-offset response from being attributed after a
|
||||||
longer horizons such as 1.2--2.0 seconds.
|
new other onset (and symmetrically censor yield after the next other
|
||||||
3. **Context length**: compare 4, 10, 15, and 20 seconds without adding an RNN.
|
offset). Freeze scenario rules without tuning on EasyCom test Sessions 1--3.
|
||||||
4. **Layer ablation**: compare audio-encoder output and several Thinker layers,
|
2. **P1 — representation diagnosis:** compare EgoSpeak features, Qwen
|
||||||
because an intermediate layer may be more discriminative than the final
|
encoder-only A/V/AV, the global Thinker state, and anonymous person slots on
|
||||||
generation-oriented layer.
|
aggregate activity and per-person active-speaker probes. Include wrong-person,
|
||||||
5. **Semantic-completion subset**: contrast completed utterances with
|
time-shifted, visual-only, and audio-only controls.
|
||||||
hesitation/incomplete pauses while matching pause duration and basic
|
3. **P2 — behavioral baselines:** evaluate `G`, `G+A`, `G+A+pooled-V`, shuffled
|
||||||
prosody as closely as possible.
|
visual features, and a parameter/history-matched temporal control using
|
||||||
6. **Question/statement/backchannel subsets**: report target-speaker AP and
|
entry/yield event AP and third-party false entries at matched onset recall.
|
||||||
premature-interruption rate on linguistically meaningful events.
|
4. **P3 — primary method:** use a shared causal person encoder plus
|
||||||
7. **Causal transcript baseline**: compare Omni audio with causal ASR text plus
|
wearer-state-conditioned attention, then ablate per-person active-speaker,
|
||||||
a language model to isolate lexical/semantic contribution.
|
addressee, and wearer event supervision.
|
||||||
|
5. **P4 — Omni adaptation ceiling:** compare parameter-matched Thinker LoRA,
|
||||||
|
local person-memory adapters, and adapter+LoRA only if person features pass
|
||||||
|
P1 and improve floor behavior in P3.
|
||||||
|
|
||||||
The frozen-Omni head is the cleanest representation test. LoRA can be reported
|
Original 0.2--2.0-second EgoSpeak AP/mAP remains a secondary comparability
|
||||||
as a second-stage ceiling, not as the only result.
|
metric. The primary paper evidence is selective behavior: correct person
|
||||||
|
alignment must reduce false wearer entry during third-party conversation while
|
||||||
|
preserving addressed response and yield. Full hypotheses, tables, stopping
|
||||||
|
rules, and the 8×L20 execution order are in
|
||||||
|
[the current experiment plan](docs/omni_speaker_floor_experiment_plan_2026-08-30.md).
|
||||||
|
|
||||||
|
Secondary semantic analyses can still include context/layer ablations,
|
||||||
|
completed-versus-hesitation subsets, question/statement/backchannel subsets,
|
||||||
|
and a causal ASR-text baseline. They support interpretation but cannot replace
|
||||||
|
the person-alignment interventions above.
|
||||||
|
|
||||||
## Transcript-derived interaction-event proxies
|
## Transcript-derived interaction-event proxies
|
||||||
|
|
||||||
@@ -2125,7 +2178,7 @@ Do not launch full EasyCom/Ego4D extraction until all of these pass:
|
|||||||
- test session IDs never enter training, validation, normalization statistics,
|
- test session IDs never enter training, validation, normalization statistics,
|
||||||
or hyperparameter selection.
|
or hyperparameter selection.
|
||||||
|
|
||||||
Once these checks pass, the first publishable experiment is:
|
Once these checks pass, the historical Stage-0 representation baseline is:
|
||||||
|
|
||||||
```text
|
```text
|
||||||
frozen Qwen2.5-Omni-3B Thinker
|
frozen Qwen2.5-Omni-3B Thinker
|
||||||
@@ -2136,4 +2189,5 @@ frozen Qwen2.5-Omni-3B Thinker
|
|||||||
+ exact EgoSpeak split and Table 2 metric
|
+ exact EgoSpeak split and Table 2 metric
|
||||||
```
|
```
|
||||||
|
|
||||||
Repository workflow verified on 2026-08-19.
|
The repository workflow was first verified on 2026-08-19. The paper question
|
||||||
|
and gated experiment plan were revised on 2026-08-30.
|
||||||
|
|||||||
104
docs/multiparty_egocentric_floor_gap_audit_2026-08-30.md
Normal file
104
docs/multiparty_egocentric_floor_gap_audit_2026-08-30.md
Normal file
@@ -0,0 +1,104 @@
|
|||||||
|
# 多方第一视角 AV wearer floor-state prediction:文献 gap 审计
|
||||||
|
|
||||||
|
日期:2026-08-30。状态:顶会与 arXiv 文献检索后的研究定位;不是实验结论。
|
||||||
|
|
||||||
|
## 1. 结论
|
||||||
|
|
||||||
|
截至核查日期,**宽泛的“多方音视频轮次预测”已经不是 research gap**,
|
||||||
|
“第一视角 when-to-speak”“在线多方 EOT”“逐人音视频 conversational graph”也都有直接先例。
|
||||||
|
|
||||||
|
仍然可信、但必须精确定义的空白是:
|
||||||
|
|
||||||
|
> 从单个第一视角、严格因果的音视频流中,利用逐人说话人 grounding,预测不可见
|
||||||
|
> camera wearer 自身的 floor-entry(OFF→ON)与 floor-yield(ON→OFF)行为,并在自然多方
|
||||||
|
> 对话中以事件级召回、误触发和延迟评测其表现。
|
||||||
|
|
||||||
|
本轮没有找到一项工作同时满足以下六点:
|
||||||
|
|
||||||
|
1. 自然多方对话,而非固定双人;
|
||||||
|
2. 单个 wearer 的第一视角输入,而非静态第三人称、多视角同步设备或每人独立传感器;
|
||||||
|
3. 原始、连续、严格因果的音视频观测;
|
||||||
|
4. 预测主体是画面中不可见的 wearer 自身;
|
||||||
|
5. 同时研究 wearer 开口与停口/让出 floor,而非只做当前活动、next speaker 或 EOT;
|
||||||
|
6. 逐人/讲话对象条件化,并用事件级 recall、false-trigger 与 latency 评价。
|
||||||
|
|
||||||
|
“没有找到同时满足”只限于本次检索范围,不能在论文中写成不受限定的绝对结论。
|
||||||
|
|
||||||
|
## 2. 关键工作的覆盖边界
|
||||||
|
|
||||||
|
| 工作 | 已覆盖 | 相对本项目缺少什么 |
|
||||||
|
|---|---|---|
|
||||||
|
| [EgoSpeak, Findings of NAACL 2025](https://aclanthology.org/2025.findings-naacl.163/) | 第一视角、在线 AV、EasyCom/Ego4D、预测 wearer 未来活动 | 将所有 others 合并为一类;三路未来逐帧分类,不显式建模逐人关系、wearer yield 或事件级误触发 |
|
||||||
|
| [3M-Transformer, ICASSP 2024](https://arxiv.org/pdf/2310.14859) | 多方、AVT、egocentric、预测 1/3/5/10s 后 next speaker | 使用至少两个参与者的同步穿戴视角及预计算文本;不是单 wearer 原始流,也不是 200ms start/yield 事件 |
|
||||||
|
| [AV-CONV, CVPR 2024](https://openaccess.thecvf.com/content/CVPR2024/html/Jia_The_Audio-Visual_Conversational_Graph_From_an_Egocentric-Exocentric_Perspective_CVPR_2024_paper.html) | 第一视角、多方、逐人 speaking/listening graph、局部头部和多通道音频 | 主要识别 clip 内各时刻的当前 conversational graph,不是严格因果的未来 wearer action forecast |
|
||||||
|
| [Online Multimodal EOT, ICMI 2024](https://graphics.cs.uh.edu/wp-content/papers/2024/2024-ICMI-OnlineTurnPrediction.pdf) | 三人、在线 EOT、overlap/interruption、文本/韵律/视线 | 受控实验室,使用 mocap、眼动仪和独立麦克风;不是自然单 wearer 第一视角原始 AV,也不以 wearer start/yield 为动作目标 |
|
||||||
|
| [Triadic VAP, Interspeech 2025](https://www.isca-archive.org/interspeech_2025/elmers25_interspeech.html) | 三方未来 voice activity projection | 纯音频、逐人通道、固定三人,不利用第一视角人物证据 |
|
||||||
|
| [MuVAP, arXiv 2026](https://arxiv.org/pdf/2606.16731) | 因果、多方、单目、单声道、face grounding、未来 floor holder | 最强 novelty threat;数据明确采用静态第三方视角,预测可见参与者的 current/next speaker,不预测不可见 wearer 的 response/yield |
|
||||||
|
| [Gaze-Enhanced Triadic Turn-Taking, Interspeech 2025](https://www.isca-archive.org/interspeech_2025/heo25_interspeech.html) | 三方、egocentric gaze、VAD、transition-level 分析 | 依赖专用眼动/空间传感器,目标为听觉辅助,不是普通第一视角 AV wearer action |
|
||||||
|
| [TurnBench, arXiv 2026](https://arxiv.org/abs/2608.25218) | 严格因果 EOT/INT,recall/FPR/latency | 双人、纯音频;适合作为事件评测范式,不是同输入/任务的直接竞争者 |
|
||||||
|
|
||||||
|
## 3. 最容易误写的新颖性声明
|
||||||
|
|
||||||
|
以下说法不能使用:
|
||||||
|
|
||||||
|
- “第一个多方 turn-taking 模型”;
|
||||||
|
- “第一个多方音视频 turn-taking 模型”;
|
||||||
|
- “第一个第一视角 when-to-speak 模型”;
|
||||||
|
- “第一个在线多方 EOT 模型”;
|
||||||
|
- “第一个逐人音视频 conversational graph”;
|
||||||
|
- “第一个 speaker-aware causal AV floor model”。
|
||||||
|
|
||||||
|
推荐使用带齐限定条件的表述:
|
||||||
|
|
||||||
|
> To our knowledge, this is the first study of strictly causal, speaker-grounded
|
||||||
|
> audiovisual forecasting of both floor-entry and floor-yield actions for an invisible
|
||||||
|
> egocentric wearer in natural multiparty conversations.
|
||||||
|
|
||||||
|
最终投稿前还应重新检索,并在正文紧接着解释 `single-egocentric-stream`、`invisible wearer`、
|
||||||
|
`both entry and yield`、`speaker-grounded` 和 `event-level evaluation` 分别排除了哪些先例。
|
||||||
|
|
||||||
|
## 4. 文献主线与必须引用的关系
|
||||||
|
|
||||||
|
### 未来活动与轮次预测
|
||||||
|
|
||||||
|
- [Voice Activity Projection, Interspeech 2022](https://www.isca-archive.org/interspeech_2022/ekstedt22_interspeech.html):
|
||||||
|
未来联合语音活动及 shift/backchannel 基础。
|
||||||
|
- [Triadic VAP, Interspeech 2025](https://www.isca-archive.org/interspeech_2025/elmers25_interspeech.html):
|
||||||
|
将 VAP 扩展到固定三人逐通道输入。
|
||||||
|
- [MuVAP, arXiv 2026](https://arxiv.org/abs/2606.16731):
|
||||||
|
单摄像机、单声道、逐人 grounding 的因果多方预测,是方法比较中必须正面处理的工作。
|
||||||
|
|
||||||
|
### 第一视角多方交互
|
||||||
|
|
||||||
|
- [3M-Transformer, ICASSP 2024](https://arxiv.org/abs/2310.14859):多视角 EgoCom 上的 next-speaker anticipation。
|
||||||
|
- [AV-CONV, CVPR 2024](https://arxiv.org/abs/2312.12870):从第一视角建立 wearer 与其他人的 speaking/listening graph。
|
||||||
|
- [EgoSpeak, Findings of NAACL 2025](https://aclanthology.org/2025.findings-naacl.163/):
|
||||||
|
与本项目最接近的数据、标签和 wearer future-activity 前身。
|
||||||
|
|
||||||
|
### 事件评测与多人 MLLM 动机
|
||||||
|
|
||||||
|
- [TurnBench](https://turnbench.sesame.com/):EOT/INT 的事件级 recall、false-positive rate 和 latency。
|
||||||
|
- [Multi-speaker Attention Alignment, CVPR 2026](https://openaccess.thecvf.com/content/CVPR2026/html/Ouyang_Multi-speaker_Attention_Alignment_for_Multimodal_Social_Interaction_CVPR_2026_paper.html):
|
||||||
|
当前 MLLM 的说话人话语—视觉区域对齐短板。
|
||||||
|
- [AMusE, CVPR 2026](https://openaccess.thecvf.com/content/CVPR2026/html/Chowdhury_AMusE_Audio-Visual_Benchmark_and_Alignment_Framework_for_Agentic_Multi-Speaker_Understanding_CVPR_2026_paper.html):
|
||||||
|
多说话人身份、角色和跨时间 grounding 仍困难,为方法动机提供证据,但不是 floor-action 任务。
|
||||||
|
- [SICNet, ICASSP 2024](https://www.amazon.science/publications/long-term-social-interaction-context-the-key-to-egocentric-addressee-detection):
|
||||||
|
第一视角 talking-to-me/addressee 已有先例,不能把“讲话对象”本身写成首次提出。
|
||||||
|
|
||||||
|
## 5. 对当前论文主线的约束
|
||||||
|
|
||||||
|
gap 本身不足以形成 ICASSP 方法贡献。论文需要证明逐人视觉不是普通容量增益:
|
||||||
|
|
||||||
|
1. 在相同 wearer onset recall 下,正确的逐人视觉减少其他人彼此交谈期间的 false entry;
|
||||||
|
2. 对真正向 wearer 交接的场景,response recall 与 latency 不退化;
|
||||||
|
3. wearer 已在说话时,对真实竞争性进入的 yield 预测改善;
|
||||||
|
4. 正确人物/时间对齐优于人物打乱、时间错配、普通全图或等预算视觉控制;
|
||||||
|
5. G、G+A、G+V、G+A+V 及 pooled/person-grounded fusion 使用相同训练和评测协议。
|
||||||
|
|
||||||
|
方法上优先考虑 speaker-grounded、wearer-state-conditioned 的 entry/yield hazard heads 或
|
||||||
|
逐人 AV 融合,而不是只增加一个普通池化旁路。TurnBench 风格事件评测应作为贡献之一,
|
||||||
|
但不能替代方法增量。
|
||||||
|
|
||||||
|
EasyCom 的逐人时间/框/讲话对象最适合验证“other→other 时不误开口”;Ego4D 可验证规模与
|
||||||
|
泛化,但相同语义分层只能在 `is_at_me`、逐人轨迹和 voice segment 确认覆盖的样本上报告。
|
||||||
|
只用 wearer 活动真值可以研究实际 start/yield,不能据此声称模型理解了 EOT、INT 或讲话意图。
|
||||||
262
docs/omni_speaker_floor_experiment_plan_2026-08-30.md
Normal file
262
docs/omni_speaker_floor_experiment_plan_2026-08-30.md
Normal file
@@ -0,0 +1,262 @@
|
|||||||
|
# Omni 多人说话者理解与 wearer floor action:实验主线(2026-08-30)
|
||||||
|
|
||||||
|
## 1. 论文真正要回答的问题
|
||||||
|
|
||||||
|
本项目不再把“Qwen2.5-Omni 是否能替换 EgoSpeak 的 ResNet、wav2vec 和
|
||||||
|
GRU/Transformer”作为最终主线。它只是一个 representation baseline。
|
||||||
|
|
||||||
|
主问题应收敛为:
|
||||||
|
|
||||||
|
> 严格因果的 Omni 模型是否真正编码了多人场景中的“谁在何时说话、话语是否面向
|
||||||
|
> wearer”这类视听 grounding;把这种逐人证据加入预测后,能否使不可见的 camera
|
||||||
|
> wearer 更合理地进入或让出话轮?
|
||||||
|
|
||||||
|
最重要的可证伪目标是:
|
||||||
|
|
||||||
|
> 在相同 wearer 开口事件召回率下,正确对齐的逐人视觉应减少 other↔other 交谈期间
|
||||||
|
> 的错误开口预测;同时,不能降低对真正 addressed-to-wearer 接话事件的召回率,也
|
||||||
|
> 不能恶化 wearer 正在讲话时的让话预测。
|
||||||
|
|
||||||
|
这比“整体 mAP 提高”更能说明模型理解了多人交互,而不只是利用音量、静音和数据
|
||||||
|
先验。
|
||||||
|
|
||||||
|
## 2. 可以和不可以主张什么
|
||||||
|
|
||||||
|
### 可以主张的研究空缺
|
||||||
|
|
||||||
|
现有工作分别研究了 egocentric 未来说话状态、多人 EOT/VAP、视听对话图和说话者
|
||||||
|
grounding,但它们尚未完整覆盖以下交集:
|
||||||
|
|
||||||
|
- 单个 egocentric wearer 视角;
|
||||||
|
- 严格因果的音频和视频;
|
||||||
|
- wearer 本人通常不在画面中,却是需要预测行为的主体;
|
||||||
|
- 同时评估进入话轮和让出话轮;
|
||||||
|
- 使用逐人/对象指向的视觉证据;
|
||||||
|
- 用事件召回、错误开口率和响应延迟,而不只用逐帧未来活动 mAP。
|
||||||
|
|
||||||
|
更稳妥的论文表述是:
|
||||||
|
|
||||||
|
> To our knowledge, this is the first study of strictly causal,
|
||||||
|
> speaker-grounded audiovisual forecasting of both floor-entry and floor-yield
|
||||||
|
> actions for an invisible egocentric wearer in natural multiparty
|
||||||
|
> conversations.
|
||||||
|
|
||||||
|
这个 claim 仍需在投稿前按检索结果逐条核对,不能写成“首个多人视听 turn-taking”或
|
||||||
|
“首个在线多人 EOT”。完整边界见
|
||||||
|
[多人 egocentric floor gap audit](multiparty_egocentric_floor_gap_audit_2026-08-30.md)。
|
||||||
|
|
||||||
|
### 不足以成为贡献的内容
|
||||||
|
|
||||||
|
- 仅把 EgoSpeak 的旧 backbone 换成 Qwen2.5-Omni;
|
||||||
|
- 仅报告 0.2--2.0 秒逐帧 mAP;
|
||||||
|
- 仅发现 G+A+V 高于 G,但没有正确人物、错位人物或参数匹配控制;
|
||||||
|
- 把 aggregate `others` 的边界直接称为语义 EOT、INT 或 addressee 真值;
|
||||||
|
- 把冻结 Omni 特征加 MLP 称为任务级端到端训练;
|
||||||
|
- 继续优化此前泛化的 transition auxiliary loss,而不区分 wearer entry/yield 场景。
|
||||||
|
|
||||||
|
冻结 Thinker 加 head 应称为 **frozen-backbone adaptation** 或 representation probe。
|
||||||
|
只有联合优化局部 adapter/LoRA 与任务目标时,才接近 task-adapted end-to-end 模型。
|
||||||
|
|
||||||
|
## 3. 四个假设
|
||||||
|
|
||||||
|
| 假设 | 问题 | 可观察证据 |
|
||||||
|
|---|---|---|
|
||||||
|
| H1 | 全局 Omni state 是否包含 aggregate floor 信息? | wearer/others 活动与未来活动 probe 优于简单先验和 encoder-only baseline |
|
||||||
|
| H2 | 全局 Omni state 是否可靠地区分画面中的不同说话人? | 匿名 person-slot active-speaker probe 明显优于随机/时序错位视觉 |
|
||||||
|
| H3 | 逐人 grounding 是否改变 wearer 的开口决策? | matched onset recall 下,other↔other 时段 false entry/min 显著下降 |
|
||||||
|
| H4 | 这种抑制是否只是让模型更保守? | addressed-to-wearer response recall/latency 和 wearer yield 不退化 |
|
||||||
|
|
||||||
|
如果 H2 不成立,论文不能宣称原始 Omni 已经具有可靠 speaker grounding;这时逐人分支
|
||||||
|
本身就是所需的适配模块。如果 H2 成立但 H3 不成立,则 representation 有说话者信息,
|
||||||
|
但它没有转化为 floor-action 收益,也不能强行讲成成功故事。
|
||||||
|
|
||||||
|
## 4. 实验顺序
|
||||||
|
|
||||||
|
### P0:先冻结评测和数据定义
|
||||||
|
|
||||||
|
在训练新模型前完成,否则后续所有数值都可能回答错问题。
|
||||||
|
|
||||||
|
1. 核实 5 Hz `row t` 的输入截止时间和 `t+h` 标签物理时间;保留现有 causal-valid
|
||||||
|
mask。
|
||||||
|
2. 修正事件归因的 competing risk:
|
||||||
|
- other offset 后的 wearer response 只在下一次 other onset 之前归因;
|
||||||
|
- other onset 后的 wearer yield 只在下一次 other offset 之前归因。
|
||||||
|
现有 `evaluate_floor_state_changes.py` 使用固定 horizon 内任意 transition,暂时只能作为
|
||||||
|
proxy,可能把后来无关的动作归给前一个边界。
|
||||||
|
3. 固定 EasyCom train/test(4--12 / 1--3);阈值、窗口和规则不能在 test 上调。
|
||||||
|
4. 从训练 session 内留出 calibration/dev 子集,用来选择阈值和 early stopping;最终表格
|
||||||
|
仍在官方 test 上一次性报告。
|
||||||
|
5. 冻结三个场景切片:
|
||||||
|
- `third_party`: other↔other,wearer 不应开口;
|
||||||
|
- `addressed`: 他人明确面向 wearer,期待 wearer 响应;
|
||||||
|
- `group_or_unknown`: 群体/目标未知,只作独立报告。
|
||||||
|
6. 审计 EasyCom `Target_of_Speech` 的覆盖与可靠性;Ego4D 缺少充分 addressee 标注时,
|
||||||
|
不把它用于 addressed-to-wearer 主指标。
|
||||||
|
|
||||||
|
P0 的产物应包括事件计数、正负样本数、示例时间轴和人工复核样例,而不只是一个 AP。
|
||||||
|
|
||||||
|
### P1:先诊断 Omni 是否真的知道“谁在说话”
|
||||||
|
|
||||||
|
使用冻结表示和轻量、容量匹配的 probe,避免先训练复杂模型后再猜收益来源。
|
||||||
|
|
||||||
|
| ID | 输入表示 | 目的 |
|
||||||
|
|---|---|---|
|
||||||
|
| D0 | EgoSpeak 原始音频/视觉特征 | 可参照的旧表示基线 |
|
||||||
|
| D1-A/V/AV | Qwen audio/visual encoder-only | 区分 encoder 信息和 Thinker 融合信息 |
|
||||||
|
| D2 | 全局 Thinker state `G` | 检查 Omni 全局状态 |
|
||||||
|
| D3 | 匿名 person slots `V_i` | 逐人 active-speaker / visible-speaker probe |
|
||||||
|
| D4 | `G + V_i` | 检查全局语义能否帮助人物归因 |
|
||||||
|
|
||||||
|
必须加入的控制:
|
||||||
|
|
||||||
|
- audio-only、visual-only、正确 AV;
|
||||||
|
- 同一时刻随机换人的 crop;
|
||||||
|
- 时间错位 crop;
|
||||||
|
- person slot permutation(模型不能依赖固定 person ID);
|
||||||
|
- 参数量和历史长度匹配的 pooled-person baseline。
|
||||||
|
|
||||||
|
P1 的通过条件不是一个固定绝对阈值,而是:正确对齐的人物表示稳定优于错人/错时控制,
|
||||||
|
且三次随机种子置信区间不完全被简单 pooled baseline 覆盖。否则不要进入昂贵的 Thinker
|
||||||
|
adapter 实验。
|
||||||
|
|
||||||
|
### P2:建立 floor-action 基线,并证明普通池化视觉的局限
|
||||||
|
|
||||||
|
已有结果应重新放入同一个事件协议:
|
||||||
|
|
||||||
|
| ID | 模型 | 已知作用 |
|
||||||
|
|---|---|---|
|
||||||
|
| B0 | `G` | 全局 Thinker baseline |
|
||||||
|
| B1 | `G + A` | 当前最主要的局部音频增益来源 |
|
||||||
|
| B2 | `G + A + pooled V` | 检查普通局部视觉的边际价值 |
|
||||||
|
| B3 | `G + A + shuffled V` | 判断视觉收益是否来自正确对齐 |
|
||||||
|
| B4 | `G + A + temporal pooled V` | 与 person-wise temporal model 做公平容量/历史控制 |
|
||||||
|
|
||||||
|
当前 EasyCom seed-0 结果中,B1 到 B2 的 0.2 秒 wearer AP 只约 `+0.02`,mean mAP
|
||||||
|
约 `+0.29`。这说明“有视觉输入”尚不等于“理解谁在对谁说话”,也正好构成逐人方法的
|
||||||
|
动机。不要把这点写成视觉已被证实有效。
|
||||||
|
|
||||||
|
逐帧 10-horizon AP/mAP 保留为与 EgoSpeak 可比的次级指标;主指标改为:
|
||||||
|
|
||||||
|
- wearer entry event AP;
|
||||||
|
- wearer yield event AP;
|
||||||
|
- matched entry recall(建议 0.7/0.8/0.9)下 third-party false entries/min;
|
||||||
|
- addressed-to-wearer response recall 和 latency;
|
||||||
|
- wearer 已经讲话时的 yield recall 和 latency。
|
||||||
|
|
||||||
|
### P3:第一优先方法——匿名逐人时序旁路
|
||||||
|
|
||||||
|
先做易解释、能在 8×L20 上完成的旁路,而不是直接修改整个 Thinker。
|
||||||
|
|
||||||
|
```text
|
||||||
|
global Thinker states G_t ───────────────────────────────┐
|
||||||
|
recent local audio A_t -> causal audio temporal encoder ─┤
|
||||||
|
person crops V_{t,i} -> shared causal encoder per person ┤
|
||||||
|
-> wearer-state-conditioned attention
|
||||||
|
-> floor entry/yield head
|
||||||
|
```
|
||||||
|
|
||||||
|
关键约束:
|
||||||
|
|
||||||
|
- 人物轨迹使用匿名 slot,共享视觉编码器和时序 GRU/TCN,避免依赖 person ID;
|
||||||
|
- 每个时刻显式 mask 不可见人物;
|
||||||
|
- causal attention 只能看到当前及过去;
|
||||||
|
- 用 wearer 当前状态门控 entry/yield 分支;
|
||||||
|
- 所有比较匹配历史长度、参数量、优化器、训练轮数与种子。
|
||||||
|
|
||||||
|
逐步消融:
|
||||||
|
|
||||||
|
| ID | 增量 | 回答的问题 |
|
||||||
|
|---|---|---|
|
||||||
|
| A0t | `G+A` + 参数匹配时序层 | 收益是否只是更多 temporal capacity? |
|
||||||
|
| A1 | pooled-person temporal branch | 收益是否只是人物区域平均更干净? |
|
||||||
|
| A2 | person-wise shared encoder + causal cross-attention | 保留人物结构是否必要? |
|
||||||
|
| A3 | A2 + per-person active-speaker auxiliary loss | 显式“谁在说”监督是否必要? |
|
||||||
|
| A4 | A3 + addressee auxiliary loss | “对谁说”是否降低 third-party false entry? |
|
||||||
|
| A5 | A4 + wearer entry/yield hazard loss | grounding 是否能转化为行为事件? |
|
||||||
|
|
||||||
|
论文的关键对照是 A2/A3/A4 对 A0t、A1、B3,而不是只对最弱的 B0。
|
||||||
|
|
||||||
|
### P4:第二阶段方法——把人物记忆接入 Omni Thinker
|
||||||
|
|
||||||
|
只有 P1 证明人物特征可解码、P3 证明人物结构对 floor action 有效后,再做:
|
||||||
|
|
||||||
|
| ID | 方法 | 作用 |
|
||||||
|
|---|---|---|
|
||||||
|
| T0 | Thinker LoRA,无人物 memory | 参数与任务适配控制 |
|
||||||
|
| T1 | local person-memory adapter,冻结 Thinker | 测试 trainable modality connection |
|
||||||
|
| T2 | person-memory adapter + upper-layer LoRA | 最接近 task-adapted end-to-end 的上限 |
|
||||||
|
|
||||||
|
adapter 应把每个 200 ms 或短 causal chunk 的人物 token 作为有 mask 的 memory,而不是把
|
||||||
|
所有人脸拼成一张图后再让模型猜人物边界。T1/T2 必须与 T0 参数匹配,并继续做 wrong
|
||||||
|
person、time-shift 和 remove-video 干预。若 T1 不优于 P3,就保留更简单、更容易解释的
|
||||||
|
旁路作为论文方法。
|
||||||
|
|
||||||
|
## 5. 主表和证据链
|
||||||
|
|
||||||
|
### 主表
|
||||||
|
|
||||||
|
每个方法至少 3 seeds,报告 mean ± sample standard deviation,并对 session/event 做
|
||||||
|
bootstrap 置信区间:
|
||||||
|
|
||||||
|
| Method | Entry event AP | Yield event AP | False entry/min @ R=.8 | Addressed recall | Response latency | EgoSpeak mean mAP |
|
||||||
|
|---|---:|---:|---:|---:|---:|---:|
|
||||||
|
|
||||||
|
其中 `False entry/min @ R=.8` 应作为最容易读懂的 headline metric;R=.7/.9 放附表或曲线。
|
||||||
|
|
||||||
|
### 三个必要的因果证据
|
||||||
|
|
||||||
|
1. **正确人物 > 错误人物**:随机换人不能保留收益。
|
||||||
|
2. **正确时间 > 错位时间**:未来帧绝不能进入输入,过去错位也应降低收益。
|
||||||
|
3. **选择性改善而非整体保守**:third-party false entry 降低时,addressed response recall
|
||||||
|
不下降。
|
||||||
|
|
||||||
|
### 定性案例
|
||||||
|
|
||||||
|
挑选固定规则而不是事后挑最好看的例子:
|
||||||
|
|
||||||
|
- 两个 visible others 连续互相讲话,wearer 保持沉默;
|
||||||
|
- visible other 转向/面向 wearer 后,wearer 在合理延迟内开口;
|
||||||
|
- wearer 正在讲话,other 竞争性进入后 wearer 让话;
|
||||||
|
- 失败案例:遮挡、背脸、多人同时讲话、目标未知。
|
||||||
|
|
||||||
|
每个案例画同一时间轴:人物框、per-person speaking score、addressee score、wearer
|
||||||
|
entry/yield score、真实 wearer 状态。
|
||||||
|
|
||||||
|
## 6. 8×L20 的执行优先级
|
||||||
|
|
||||||
|
不要一开始把 8 张卡都投入 T2。建议按 gate 推进:
|
||||||
|
|
||||||
|
1. **第 1 轮**:3 seeds 的 B1、B2、A0t,加 B3 wrong-alignment control;冻结评测。
|
||||||
|
2. **第 2 轮**:D3/D4 person probe;确认人物和时间对齐真的可解码。
|
||||||
|
3. **第 3 轮**:3 seeds 的 A1、A2、A3;根据 matched-recall 主指标选方向。
|
||||||
|
4. **第 4 轮**:只对通过的 A3 增加 addressee/event supervision。
|
||||||
|
5. **第 5 轮**:T0/T1;只有 T1 显著优于旁路后才做 T2。
|
||||||
|
6. **跨数据集**:EasyCom 用于完整 speaker/addressee 分析;Ego4D 用于 activity、entry/yield
|
||||||
|
泛化,不强行报告其不支持的 addressee 主指标。
|
||||||
|
|
||||||
|
每轮都先在一个 session 做 shape、时间和 leakage smoke test,再并行 full run。模型选择只看
|
||||||
|
train 内 calibration/dev,不读取官方 test 结果做超参数搜索。
|
||||||
|
|
||||||
|
## 7. 结果如何决定论文方向
|
||||||
|
|
||||||
|
| 观察 | 应得结论 |
|
||||||
|
|---|---|
|
||||||
|
| 正确视觉与错位视觉无差别 | 当前局部视觉没有被模型使用;不能讲 AV social grounding |
|
||||||
|
| 只有 audio 提升 | 收敛成 prosody/audio floor forecasting,不硬写视觉故事 |
|
||||||
|
| person probe 好,但 floor 指标无提升 | Omni 有人物信息,但当前任务/融合没有利用它 |
|
||||||
|
| A2 优于 pooled,但 A3/A4 无额外收益 | 人物结构有效,辅助标签质量或目标设计需调整 |
|
||||||
|
| A4 在 matched recall 下降低 third-party false entry,addressed recall 不退化 | 最强、最直接的多人 AV floor-action 证据 |
|
||||||
|
| T1/T2 不优于旁路 | 采用简单旁路;不为“端到端”牺牲可解释性 |
|
||||||
|
|
||||||
|
## 8. 最小可投稿版本与强版本
|
||||||
|
|
||||||
|
**最小可投稿版本**:新的 wearer entry/yield 事件协议、Omni speaker-grounding 诊断、
|
||||||
|
EgoSpeak/encoder-only/Thinker/local-audio/pooled-video 强基线、EasyCom+Ego4D 复现,并清楚
|
||||||
|
展示逐帧 mAP 隐藏的行为差异。
|
||||||
|
|
||||||
|
**强版本**:在上述基础上,匿名逐人 causal fusion 在 matched wearer recall 下显著降低
|
||||||
|
third-party false entry,保持 addressed response 与 yield,并通过 wrong-person、time-shift、
|
||||||
|
audio-only、参数匹配和跨数据集控制。
|
||||||
|
|
||||||
|
前者回答“现有 Omni 到底会不会管理话轮”;后者进一步给出“怎样让它利用多人视觉管理
|
||||||
|
话轮”。后者才是更有竞争力的 ICASSP 主线。
|
||||||
@@ -2,12 +2,18 @@
|
|||||||
|
|
||||||
- 作者:Mehdi Fatan Serj、Emanuele Mincato、Dimitra Pintzou、Mariella Dimiccoli
|
- 作者:Mehdi Fatan Serj、Emanuele Mincato、Dimitra Pintzou、Mariella Dimiccoli
|
||||||
- 发表:ICASSP 2024,8050–8054
|
- 发表:ICASSP 2024,8050–8054
|
||||||
- 阅读深度:作者机构摘要及正式会议信息筛选
|
- 阅读深度:方法、数据、输入和预测协议重点阅读
|
||||||
- 记录日期:2026-08-28
|
- 记录日期:2026-08-30
|
||||||
- 来源:[作者机构](https://www.iri.upc.edu/publications/show/2887)、[DOI](https://doi.org/10.1109/ICASSP48485.2024.10448136)
|
- 来源:[论文](https://arxiv.org/pdf/2310.14859)、[作者机构](https://www.iri.upc.edu/publications/show/2887)、[DOI](https://doi.org/10.1109/ICASSP48485.2024.10448136)
|
||||||
|
|
||||||
## 摘要与项目关系
|
## 摘要与项目关系
|
||||||
|
|
||||||
在 EgoCom 的同步、多视角 embodied 对话数据上,通过多阶段、多流 multimodal transformer 预测轮次。强调不同人物视角和模态间的交互。
|
在 EgoCom 的同步、多视角 embodied 对话数据上,通过多阶段、多流 multimodal transformer
|
||||||
|
融合 audio/video/text,预测 1、3、5、10 秒后的 next speaker。EgoCom 包含三人对话,至少
|
||||||
|
两人佩戴同步音视频眼镜;输入使用预计算视觉、说话人音频和 transcript embeddings,并聚合
|
||||||
|
过去 4、5、10 或 30 秒。
|
||||||
|
|
||||||
这是多人时序融合的相关工作,但 EgoCom 不等于 EasyCom,多个参与者摄像机也不同于单 wearer 视角。本轮未完整核对方法、输入因果性及代码,不能直接作为可复现命令来源。
|
这是“多方+第一视角+AVT+未来 next speaker”的直接先例。它与本项目的区别不能只写数据集:
|
||||||
|
其系统依赖多位参与者的同步穿戴视角和已有文本,预测远期 speaker class;本项目目标是单个
|
||||||
|
wearer 传感流下不可见 wearer 的 200ms 级 entry/yield 事件及误触发。论文表格中的高 accuracy
|
||||||
|
还混合了 `Use Prior=True` 条件,不能与本项目逐帧 AP 或事件指标直接比较。
|
||||||
|
|||||||
18
literature_notes/50_av_conversational_graph.md
Normal file
18
literature_notes/50_av_conversational_graph.md
Normal file
@@ -0,0 +1,18 @@
|
|||||||
|
# The Audio-Visual Conversational Graph: From an Egocentric-Exocentric Perspective
|
||||||
|
|
||||||
|
- 作者:Wenqi Jia、Miao Liu、Hao Jiang、Ishwarya Ananthabhotla、James M. Rehg、Vamsi Krishna Ithapu、Ruohan Gao
|
||||||
|
- 发表:CVPR 2024,26396–26405
|
||||||
|
- 阅读深度:摘要、问题定义、输入、标注及补充实验重点阅读
|
||||||
|
- 核查日期:2026-08-30
|
||||||
|
- 来源:[CVF](https://openaccess.thecvf.com/content/CVPR2024/html/Jia_The_Audio-Visual_Conversational_Graph_From_an_Egocentric-Exocentric_Perspective_CVPR_2024_paper.html)、[arXiv](https://arxiv.org/abs/2312.12870)
|
||||||
|
|
||||||
|
## 摘要与项目关系
|
||||||
|
|
||||||
|
从第一视角音视频构造动态 conversational graph,同时预测 camera wearer 与画面中各人物的
|
||||||
|
speaking/listening 关系。模型显式融合跨时间、跨人物和跨模态信息,并使用人物头部 crop、
|
||||||
|
空间 mask 与多通道音频。实验数据包含多说话人及并行会话。
|
||||||
|
|
||||||
|
这是“第一视角+多方+逐人 AV 交互图”的直接先例,因此不能把这些组件本身作为首次贡献。
|
||||||
|
但其任务是根据约 3s clip 重建各时刻当前的 speaking/listening graph,补充材料说明各帧结果
|
||||||
|
一次性预测;没有建立严格因果的 wearer 未来 entry/yield 事件协议。其多通道音频、提供的头框
|
||||||
|
以及预定义交谈组也不同于本项目的单 wearer 流式输入。
|
||||||
18
literature_notes/51_online_multiparty_eot.md
Normal file
18
literature_notes/51_online_multiparty_eot.md
Normal file
@@ -0,0 +1,18 @@
|
|||||||
|
# Online Multimodal End-of-Turn Prediction for Three-party Conversations
|
||||||
|
|
||||||
|
- 作者:Meng-Chen Lee、Zhigang Deng
|
||||||
|
- 发表:ICMI 2024,57–65;Best Paper Runner-up
|
||||||
|
- 阅读深度:任务、采集、方法、评测及局限重点阅读
|
||||||
|
- 核查日期:2026-08-30
|
||||||
|
- 来源:[作者论文 PDF](https://graphics.cs.uh.edu/wp-content/papers/2024/2024-ICMI-OnlineTurnPrediction.pdf)、[DOI](https://doi.org/10.1145/3678957.3685742)
|
||||||
|
|
||||||
|
## 摘要与项目关系
|
||||||
|
|
||||||
|
用连续窗口及 PLM-GRU 融合文本、韵律、视线和交互状态,在线预测三方对话中的 end-of-turn,
|
||||||
|
标注包含 turn-taking、interruption 与 overlap。相对依赖停顿后的 IPU 方法,它强调连续预测
|
||||||
|
和低延迟,因此“第一个在线多方 EOT”不能作为本项目 claim。
|
||||||
|
|
||||||
|
数据来自受控实验室:八台光学 mocap 相机、每人眼动仪、mocap suit 和独立无线麦克风。
|
||||||
|
输入是工程化 gaze/prosody/text/state,而不是自然环境中的单个第一视角原始 AV;任务检测当前
|
||||||
|
说话人的 EOT,也不等于预测不可见 wearer 自身何时进入或退出 floor。论文还承认尚未形成
|
||||||
|
完整的实时多模态 top-down 系统。
|
||||||
18
literature_notes/52_turnbench.md
Normal file
18
literature_notes/52_turnbench.md
Normal file
@@ -0,0 +1,18 @@
|
|||||||
|
# TurnBench: A Multi-Domain Benchmark for Turn-Taking Dynamics in Spoken Dialogue
|
||||||
|
|
||||||
|
- 作者:Freeman Jiang 等
|
||||||
|
- 状态:2026 arXiv 预印本及公开 benchmark
|
||||||
|
- 阅读深度:摘要、任务定义、提交因果性和公开 leaderboard 重点阅读
|
||||||
|
- 核查日期:2026-08-30
|
||||||
|
- 来源:[论文](https://arxiv.org/abs/2608.25218)、[benchmark](https://turnbench.sesame.com/)、[官方仓库](https://github.com/SesameAILabs/turnbench)
|
||||||
|
|
||||||
|
## 摘要与项目关系
|
||||||
|
|
||||||
|
建立跨六类对话的双人双通道 EOT/INT benchmark。模型必须提交其作出决定的时间,而不是事后
|
||||||
|
回填事件发生时间;主要指标是固定工作点下的 recall、false-positive rate 和 latency。
|
||||||
|
结果显示 backchannel 密集的对话特别容易造成 interruption false positives,而且现有系统尚未
|
||||||
|
同时达到快速、高召回和低误报。
|
||||||
|
|
||||||
|
它不是多方、音视频或第一视角任务,但提供了比逐帧 AP 更接近 agent 行为的评测依据。
|
||||||
|
本项目应借鉴一对一事件匹配、因果提交时间、匹配召回率和误报预算,同时保留 wearer entry/yield
|
||||||
|
及多方场景分层;不能直接把由二值活动自动派生的 proxy 称为 TurnBench 人工 EOT/INT 真值。
|
||||||
16
literature_notes/53_amuse.md
Normal file
16
literature_notes/53_amuse.md
Normal file
@@ -0,0 +1,16 @@
|
|||||||
|
# AMusE: Audio-Visual Benchmark and Alignment Framework for Agentic Multi-Speaker Understanding
|
||||||
|
|
||||||
|
- 作者:Sanjoy Chowdhury 等
|
||||||
|
- 发表:CVPR 2026,22998–23009
|
||||||
|
- 阅读深度:摘要及官方论文页筛选
|
||||||
|
- 核查日期:2026-08-30
|
||||||
|
- 来源:[CVF](https://openaccess.thecvf.com/content/CVPR2026/html/Chowdhury_AMusE_Audio-Visual_Benchmark_and_Alignment_Framework_for_Agentic_Multi-Speaker_Understanding_CVPR_2026_paper.html)
|
||||||
|
|
||||||
|
## 摘要与项目关系
|
||||||
|
|
||||||
|
系统评估 MLLM 在多说话人对话中的人物身份追踪、角色保持及音视频事件时序 grounding,报告
|
||||||
|
当前通用模型在 agentic multi-speaker understanding 上仍有明显短板。
|
||||||
|
|
||||||
|
该工作不是 wearer floor-entry/yield 预测,也不能作为直接数值基线;它适合与 Multi-speaker
|
||||||
|
Attention Alignment 一起支撑“简单加入视觉并不等于获得逐人社会理解”的动机。引用时应把
|
||||||
|
其 benchmark/理解任务与本项目 200ms 因果行为预测明确分开。
|
||||||
@@ -2,9 +2,11 @@
|
|||||||
|
|
||||||
本目录持久化记录我们在本项目讨论中实际阅读、分析或筛选过的论文。这里的“摘要”均为面向当前课题的中文转述,不是原文摘要的复制。
|
本目录持久化记录我们在本项目讨论中实际阅读、分析或筛选过的论文。这里的“摘要”均为面向当前课题的中文转述,不是原文摘要的复制。
|
||||||
|
|
||||||
最后更新:2026-08-28
|
最后更新:2026-08-30
|
||||||
|
|
||||||
最新专题:[可训练多人 AV 融合、8×L20 与 matched-recall 评测](../docs/trainable_speaker_fusion_survey_2026-08-28.md)。新增 33–49:涵盖逐人时序融合、讲话对象、模型内适配器、模态竞争及专门 AV 编码器,区分正式发表、作者接收声明和未发布代码。计划尚未实施;文献已发表不代表迁移到本项目已有效。
|
最新专题:[多方第一视角 AV wearer floor-state prediction 文献 gap 审计](../docs/multiparty_egocentric_floor_gap_audit_2026-08-30.md)。结论是宽泛组件均有先例;仍可信的交集是单 wearer 严格因果流、不可见 wearer 的 entry/yield、逐人 grounding 和事件级评测。MuVAP、EgoSpeak、3M-Transformer、AV-CONV 与在线多方 EOT 必须正面区分。
|
||||||
|
|
||||||
|
此前专题:[可训练多人 AV 融合、8×L20 与 matched-recall 评测](../docs/trainable_speaker_fusion_survey_2026-08-28.md)。新增 33–49:涵盖逐人时序融合、讲话对象、模型内适配器、模态竞争及专门 AV 编码器,区分正式发表、作者接收声明和未发布代码。计划尚未实施;文献已发表不代表迁移到本项目已有效。
|
||||||
|
|
||||||
同日此前专题:[逐人 chunk、train-free Thinker 与 wearer onset/offset 计划](../docs/speaker_aware_thinker_plan_2026-08-28.md)。新增 28–32 五篇笔记,补读 16、23,核查 24 的实现边界。
|
同日此前专题:[逐人 chunk、train-free Thinker 与 wearer onset/offset 计划](../docs/speaker_aware_thinker_plan_2026-08-28.md)。新增 28–32 五篇笔记,补读 16、23,核查 24 的实现边界。
|
||||||
|
|
||||||
@@ -69,9 +71,20 @@
|
|||||||
| 47 | [3M-Transformer](47_3m_transformer.md) | 2024 | 摘要级筛选 | ICASSP;多视角 EgoCom,区别于单视角 EasyCom |
|
| 47 | [3M-Transformer](47_3m_transformer.md) | 2024 | 摘要级筛选 | ICASSP;多视角 EgoCom,区别于单视角 EasyCom |
|
||||||
| 48 | [VA-related Pretrained Encoders for MM-VAP](48_va_pretrained_multimodal_vap.md) | 2026 | 重点章节阅读 | 作者声明 RO-MAN 接收;LoRA 与窗口因果性审计 |
|
| 48 | [VA-related Pretrained Encoders for MM-VAP](48_va_pretrained_multimodal_vap.md) | 2026 | 重点章节阅读 | 作者声明 RO-MAN 接收;LoRA 与窗口因果性审计 |
|
||||||
| 49 | [Modality Drop-Out for DDSD](49_modality_dropout_ddsd.md) | 2024 | 摘要及会议信息筛选 | ICASSP;缺失模态及固定工作点误接受率 |
|
| 49 | [Modality Drop-Out for DDSD](49_modality_dropout_ddsd.md) | 2024 | 摘要及会议信息筛选 | ICASSP;缺失模态及固定工作点误接受率 |
|
||||||
|
| 50 | [Audio-Visual Conversational Graph / AV-CONV](50_av_conversational_graph.md) | 2024 | 问题、输入和标注重点阅读 | CVPR;第一视角逐人 speaking/listening graph,但不是因果未来 wearer action |
|
||||||
|
| 51 | [Online Multimodal End-of-Turn Prediction](51_online_multiparty_eot.md) | 2024 | 任务、采集、方法和局限重点阅读 | ICMI;在线三方 EOT,但依赖实验室 gaze/mocap/独立麦克风 |
|
||||||
|
| 52 | [TurnBench](52_turnbench.md) | 2026 | 任务、因果提交和 leaderboard 重点阅读 | 双人音频事件评测;recall/FPR/latency 范式 |
|
||||||
|
| 53 | [AMusE](53_amuse.md) | 2026 | 摘要及官方论文页筛选 | CVPR;MLLM 多说话人身份、角色和 AV grounding 短板 |
|
||||||
|
|
||||||
## 当前综合判断
|
## 当前综合判断
|
||||||
|
|
||||||
|
2026-08-30 更新:宽泛的“多方 AV turn-taking”已被 MuVAP、3M-Transformer、在线多方 EOT
|
||||||
|
等工作覆盖;“第一视角 AV when-to-speak”已有 EgoSpeak,“逐人第一视角 speaking/listening”
|
||||||
|
已有 AV-CONV。当前可辩护定位是:**单个严格因果第一视角流中,利用 speaker-grounded AV
|
||||||
|
预测不可见 wearer 自身的 floor entry 与 yield,并作事件级误触发/召回/延迟评价**。
|
||||||
|
参见[完整 gap 审计](../docs/multiparty_egocentric_floor_gap_audit_2026-08-30.md)。gap 不替代方法贡献;
|
||||||
|
仍需证明逐人视觉在同 onset recall 下减少第三方交谈期间 false entry,并不损害真实 response。
|
||||||
|
|
||||||
2026-08-28 更新:[encoder-only 完整结果](../results/easycom_encoder_only_1s_4ep_2026-08-28/README.md)显示近期音频与 Thinker 历史互补;普通视觉池化对 wearer 的增益尚不清楚。先补 wearer 专属事件及讲话对象分层,再对照逐人时序旁路和模型内可训练 adapter/LoRA。现有 R1/R2 是全活动变化,不是 wearer onset/offset 的直接评测。MuVAP 等工作使“逐人+因果融合”不再足以单独构成创新;新假设是对不可见 wearer 的讲话对象条件化预测与受召回约束的第三方误开口降低。
|
2026-08-28 更新:[encoder-only 完整结果](../results/easycom_encoder_only_1s_4ep_2026-08-28/README.md)显示近期音频与 Thinker 历史互补;普通视觉池化对 wearer 的增益尚不清楚。先补 wearer 专属事件及讲话对象分层,再对照逐人时序旁路和模型内可训练 adapter/LoRA。现有 R1/R2 是全活动变化,不是 wearer onset/offset 的直接评测。MuVAP 等工作使“逐人+因果融合”不再足以单独构成创新;新假设是对不可见 wearer 的讲话对象条件化预测与受召回约束的第三方误开口降低。
|
||||||
|
|
||||||
以下为 2026-08-27 的阶段判断:
|
以下为 2026-08-27 的阶段判断:
|
||||||
|
|||||||
Reference in New Issue
Block a user