dataset

• 兩張圖：
• 注意一個區(qū)分：trimmed and untrimmed videos

outline

• action recognition
• action temporal localization
• action spatial detection
• action spatial-temporal detection

opportunities and challenges

• opportunities ：
  • videos provide huge and rich data for visual learning
  • action is core in motion perception and has many applications in video understanding
• challenges：
  • complex dynamics and temporal variations
  • action vocabulary is not well defined
  • noisy and weakly labels (dense labeling is expensive)
  • High computational and memory cost

temporal structure: 需要對動作進行分解：decomposition

常用的 Deep networks

• large-scale video classification with CNN (feifei li, CVPR2014)
• Two-Stream CNN for action recognition in videos (NIPS2014)
• learning spatiotemporal features with 3D CNN (Du Tran, ICCV2015)
• TDD (liming wang, CVPR2015)
• Real-time action recognition with enhanced motion vector CNNs (CVPR2016)
• Two Stream I3D (CVPR2017)
• R(2+1)D (CVPR2018)
• SlowFast Networks (kaiming he, CVPR2019)

liming wang 自己的3篇工作

• shot-term -> middle-term -> long term modeling，對應(yīng)的論文是 (ARTNet -> TSN -> UntrimmedNet)
• 更多細節(jié)理解，直接看他的PPT寫讀后感
• 按照 liming wang 自己說的，video action recognition/detection 對于 我的VAD 基本沒有幫助

之前看到一些很好的zhihu link: 動作識別-1, 動作識別-2, 時序行為檢測-1, 時序行為檢測-2, 時序行為檢測-3, 時序行為檢測-4,

所有的PPT圖片

DL的一些經(jīng)驗性Trick介紹

人臉識別的開集特點 (Open-set 和 novelty detection有點類似，參考TODO)

• Center Loss (ECCV2016)

  center loss意思即為：為每一個類別提供一個類別中心，最小化min-batch中每個樣本與對應(yīng)類別中心的距離，這樣就可以達到縮小類內(nèi)距離的目的。

  center loss的原理主要是在softmax loss的基礎(chǔ)上，通過對訓練集的每個類別在特征空間分別維護一個類中心，在訓練過程，增加樣本經(jīng)過網(wǎng)絡(luò)映射后在特征空間與類中心的距離約束，從而兼顧了類內(nèi)聚合與類間分離。

  From：link-1, link-2

• Center Loss的改進 (IJCV2019): 用投影方向代替類中心

• Large Margin思想設(shè)計的Loss:

  • L-softmax (Liu, ICML 2016)
  • A-softmax (Liu, CVPR2017)
  • Additive Margin Softmax (ICLR 2018 workshop)
  • CosLoss (wang, CVPR2018)
  • ArcFace (CVPR2019)

Range Loss ：有效應(yīng)對類間樣本數(shù)不均衡造成的長尾問題

• motivation：少數(shù)人(明星)具有大量的圖片，多數(shù)人卻只有少量圖片，這種長尾分布啟發(fā)了兩個動機：(1)長尾分布如何影響模型性能的理論分析；(2)設(shè)計新的Loss解決這個問題
• 此處有一張圖片，并且Range Loss 的 PPT缺失了

video action recognition

• 姿態(tài)注意力機制 RPAN (ICCV2017, Oral)
  • 把行為識別和姿態(tài)估計兩個任務(wù)進行結(jié)合
  • 利用姿態(tài)變化，引導 RNN 對行為的動態(tài)過程進行建模

一篇文章

• Temporal Hallucinating for Action Recognition with Few Still Images (CVPR 2018)

一些圖

applicaltion

• web video search
• video recommendation system
• avoid inappropriate advertisement

Tasks of Emotions in videos

• Emotion recognition
• emotion attribution
• emotion-oriented summarization

Challenges

• Sparsely expressed in videos
• Diverse content and variable quality

Knowledge Transfer

• Zero-shot Emotion learning (配一張圖)
  • A multi-task neural approach for emotion attribution, classification and summarization (TMM)
  • Frame-Transfermer emotion classification Network (*CMR 2017)

Emotion-oriented summarization

• 相當于選擇關(guān)鍵幀以及幀信息融合

Face emotion

• Posture, Expression, Identity in faces

一些圖：

outline

• introduction
• structured feature learning
• back-bone model design
• conclusion

introduction

• object detection
• human pose estimation
• action recognition

structured feature learning

• structure in neurons
  • motivation：傳統(tǒng) neurons 在同一層沒有連接，在相鄰層存在局部或者全部連接，沒有保證局部區(qū)域的信息。從而引出每一層網(wǎng)絡(luò)的各神經(jīng)元具有結(jié)構(gòu)化信息的。然后以人體姿態(tài)估計為例，分析了基于全連接神經(jīng)網(wǎng)絡(luò)的問題：在對人體節(jié)點的距離進行建模需要大的卷積核以及一些關(guān)節(jié)點的關(guān)系是不穩(wěn)定。提出結(jié)構(gòu)化特征學習的人體姿態(tài)估計模型（Bidirectional Tree）。
  • Bidirectional Tree
  • 對應(yīng)的papers
    • end-to-end learning of deformable mixture of parts and deep convolutional neural networks for human pose estimation (CVPR2016)
    • structure feature learning for pose estimation (CVPR2016)
    • CRF-CNN, modeling structured information in human pose estimation (CVPR2016)
    • learning deep structured multi-scale features using attention-gated CRFs for contour prediction (NIPS 2017)
  • application of structured feature learning
    • 有一張圖片

back-bone model design

• Hourglass for classification (Encode-Decoder 結(jié)構(gòu)，比如 UNet，一般用于圖像分割，不用于分類)
  • 希望: feature with high-level semantics and high resolution is good
  • 現(xiàn)實：feature with high-level semantics with low resolution
  • Hourglass for classification has poor performance的原因分析：Different tasks require different resolution of features，所以提出 FishNet
• FishNet
  • motivation: 為了統(tǒng)一利用像素級、區(qū)域級以及圖像級任務(wù)的優(yōu)勢，歐陽萬里老師提出了FishNet，FishNet的優(yōu)勢是：更好的將梯度傳到淺層網(wǎng)絡(luò)，所提取的特征包含了豐富的低層和高層語義信息并保留和微調(diào)了各層級信息。
  • pros.
    • better gradient flow to shallow layers
    • features:
      • contain rich low-level and high-level semantics
      • are preserved and refined from each other (信息互相交流)
  • code: https://github.com/kevin-ssy/FishNet

conclusion

• structured deep learning is (1)effective (2)from observation
• end-to-end joint training bridges the gap between structure modeling and feature learning

一些圖

• 基于軌跡的行為分析
• 面向任意視頻的行為識別 (liming wang)
• 面向監(jiān)控視頻的行為識別

• Tiny DSOD (BMVC 2018)
• Toward accurate one-stage object detection with AP-Loss (CVPR 2019)
• kill two birds with one stone: boosting both object detection accuracy and speed with adaptive patch-of-interest composition (2017)

• 三維目標檢測與姿態(tài)估計
• 多目標跟蹤
• 基于目標檢測、跟蹤的實時場景統(tǒng)計分析
• 多相機跟蹤
  • (correspondence structure Re-ID)
    • learning correspondence structure for person re-identification (TIP2017)
    • person re-identification with correspondence sturcture learning (ICCV 2015)
  • (Group Re-ID)
    • Group re-identification: leveraging and integrating multi-grain information, (MM2018)
  • 車載跨相機定位
  • 無人超市
  • 野生東北虎 Re-ID

• 多尺度特征
  • action recognition with coarse-to-fine deep feature interation and asynchronous fusion (AAAI 2018)
• 時空異步guanlian
  • cross-stream selective networks for action recognition (CVPR workshop 2019)
• 時空行為定位
  • finding action tubes with an integrated sparse-to-dense framework (arxiv 2019)
• 監(jiān)控行為識別
  • 有一張圖

• 實時行為/事件檢測
• 基于軌跡的行為分析-個體行為分析
  • a tube-and-droplet-based apporach for representing and analyzing motion trajectories (TPAMI 2017)，非DL，無好感
• 基于軌跡的行為分析-軌跡聚類與挖掘
  • unsupervised trajectory clustering via adaptive multi-kernel-based shrinkage (ICCV 2015)，比較老。。。但可以以它為base，看最新的引用它的高質(zhì)量論文即可
• 密集場景行為分析
  • a diffusion and clustering-based approach for finding coherent motions and understanding crowd scenes (TIP 2016)
  • finding coherent motions and semantic regions in crowd scenes: a diffusion and clustering apporach (ECCV 2014)