Action Recognition. Computer Vision Jia-Bin Huang, Virginia Tech. Many slides from D. Hoiem

Transcription

1 Action Recognition Computer Vision Jia-Bin Huang, Virginia Tech Many slides from D. Hoiem

2 This section: advanced topics Convolutional neural networks in vision Action recognition Vision and Language 3D Scenes and Context

3 What is an action? Action: a transition from one state to another Who is the actor? How is the state of the actor changing? What (if anything) is being acted on? How is that thing changing? What is the purpose of the action (if any)?

4 How do we represent actions? Categories Walking, hammering, dancing, skiing, sitting down, standing up, jumping Poses Nouns and Predicates <man, swings, hammer> <man, hits, nail, w/ hammer>

5 What is the purpose of action recognition? To describe

6 What is the purpose of action recognition? To predict

7 What is the purpose of action recognition? To understand the intention and motivation Predicting Motivations of Actions by Leveraging Text, CVPR 2016

8 How can we identify actions? Motion Pose Held Objects Nearby Objects

9 Representing Motion Optical Flow with Motion History Bobick Davis 2001

10 Representing Motion Space-Time Volumes Blank et al. 2005

11 Representing Motion Optical Flow with Split Channels optical flow split into pos/neg channels blurred pos/neg flow Efros et al. 2003

12 Representing Motion Tracked Points Matikainen et al. 2009

13 Representing Motion Space-Time Interest Points Moving corner Ball hits wall Corner detectors in space-time Balls collide Balls collide (different scale) Laptev 2005

14 Representing Motion Space-Time Interest Points Laptev 2005

15 Examples of Action Recognition Systems Feature-based classification Recognition using pose and objects

16 Action recognition as classification Retrieving actions in movies, Laptev and Perez, 2007

17 Remember image categorization Training Images Training Image Features Training Labels Classifier Training Trained Classifier Test Image Image Features Testing Trained Classifier Prediction Outdoor

18 Remember spatial pyramids. Compute histogram in each spatial bin

19 Features for Classifying Actions 1. Spatio-temporal pyramids Image Gradients Optical Flow

20 Features for Classifying Actions 2. Spatio-temporal interest points Corner detectors in space-time Descriptors based on Gaussian derivative filters over x, y, time

21 Classification Boosted stubs for pyramids of optical flow, gradient Nearest neighbor for STIP

22 Searching the video for an action 1. Detect keyframes using a trained HOG detector in each frame 2. Classify detected keyframes as positive (e.g., drinking ) or negative ( other )

23 Accuracy in searching video With keyframe detection Without keyframe detection

24 Talk on phone Get out of car Learning realistic human actions from movies, Laptev et al. 2008

25 Approach Space-time interest point detectors Descriptors HOG, HOF Pyramid histograms (3x3x2) SVMs with Chi-Squared Kernel Interest Points Spatio-Temporal Binning

26 Results

27 Action Recognition using Pose and Objects Modeling Mutual Context of Object and Human Pose in Human-Object Interaction Activities, B. Yao and Li Fei-Fei, 2010 Slide Credit: Yao/Fei-Fei

28 Human-Object Interaction Holistic image based classification Integrated reasoning Human pose estimation Head Torso Slide Credit: Yao/Fei-Fei

29 Human-Object Interaction Holistic image based classification Integrated reasoning Human pose estimation Object detection Tennis racket Slide Credit: Yao/Fei-Fei

30 Human-Object Interaction Holistic image based classification Integrated reasoning Human pose estimation Object detection Action categorization Tennis racket Head Torso Activity: Tennis Forehand Slide Credit: Yao/Fei-Fei

31 Human pose estimation & Object detection Human pose estimation is challenging. Difficult part appearance Self-occlusion Image region looks like a body part Felzenszwalb & Huttenlocher, 2005 Ren et al, 2005 Ramanan, 2006 Ferrari et al, 2008 Yang & Mori, 2008 Andriluka et al, 2009 Eichner & Ferrari, 2009 Slide Credit: Yao/Fei-Fei

32 Human pose estimation & Object detection Human pose estimation is challenging. Felzenszwalb & Huttenlocher, 2005 Ren et al, 2005 Ramanan, 2006 Ferrari et al, 2008 Yang & Mori, 2008 Andriluka et al, 2009 Eichner & Ferrari, 2009 Slide Credit: Yao/Fei-Fei

33 Human pose estimation & Object detection Facilitate Given the object is detected. Slide Credit: Yao/Fei-Fei

34 Human pose estimation & Object detection Small, low-resolution, partially occluded Object detection is challenging Image region similar to detection target Viola & Jones, 2001 Lampert et al, 2008 Divvala et al, 2009 Vedaldi et al, 2009 Slide Credit: Yao/Fei-Fei

35 Human pose estimation & Object detection Object detection is challenging Viola & Jones, 2001 Lampert et al, 2008 Divvala et al, 2009 Vedaldi et al, 2009 Slide Credit: Yao/Fei-Fei

36 Human pose estimation & Object detection Facilitate Given the pose is estimated. Slide Credit: Yao/Fei-Fei

37 Human pose estimation & Object detection Mutual Context Slide Credit: Yao/Fei-Fei

38 Mutual Context Model Representation A: Activity A O: Tennis forehand Croquet shot Volleyball smash Object Human pose H H: Tennis racket Croquet mallet Volleyball O P 1 P 2 Body parts P N f O f1 f2 fn Intra-class variations More than one H for each A; Unobserved during training. Image evidence P: l P : location; θ P : orientation; s P : scale. f: Shape context. [Belongie et al, 2002] Slide Credit: Yao/Fei-Fei

39 Learning Results Cricket defensive shot Cricket bowling Croquet shot Slide Credit: Yao/Fei-Fei

40 Learning Results Tennis forehand Tennis serve Volleyball smash Slide Credit: Yao/Fei-Fei

41 Model Inference I The learned models Slide Credit: Yao/Fei-Fei

42 Model Inference I The learned models Head detection Compositional Inference Torso detection Layout of the object and body parts. * * A1, H1, O1, P1, n n [Chen et al, 2007] Tennis racket detection Slide Credit: Yao/Fei-Fei

43 Model Inference I The learned models Output * * A1 H1 O1 P1, * * A, H, O, P,,,, n n K K K K n n Slide Credit: Yao/Fei-Fei

44 Dataset and Experiment Setup Sport data set: 6 classes 180 training (supervised with object and part locations) & 120 testing images Tasks: Object detection; Pose estimation; Cricket defensive shot Cricket bowling Croquet shot Activity classification. Tennis forehand Tennis serve Volleyball smash [Gupta et al, 2009] Slide Credit: Yao/Fei-Fei

45 Dataset and Experiment Setup Sport data set: 6 classes 180 training (supervised with object and part locations) & 120 testing images Tasks: Object detection; Pose estimation; Cricket defensive shot Cricket bowling Croquet shot Activity classification. Tennis forehand Tennis serve Volleyball smash [Gupta et al, 2009] Slide Credit: Yao/Fei-Fei

46 Object Detection Results Valid region Cricket bat Cricket ball Sliding window [Andriluka et al, 2009] Pedestrian context [Dalal & Triggs, 2006] Our Method Precision Recall Croquet mallet Tennis racket Volleyball Precision Recall Slide Credit: Yao/Fei-Fei

47 Precision Background clutter Small object Object Detection Results 1 Our Method 1 Our 1Method 0.8 Pedestrian as context Scanning 0.8 window detector Pedestrian as context Our Method Sliding window Pedestrian context Scanning Our window method detector 0.8 Pedestrian as context Scanning window detector Recall Recall Recall Precision Precision Precision Recall Cricket ball Volleyball Precision Recall 59 Slide Credit: Yao/Fei-Fei

48 Dataset and Experiment Setup Sport data set: 6 classes 180 training & 120 testing images Tasks: Object detection; Pose estimation; Cricket defensive shot Cricket bowling Croquet shot Activity classification. Tennis forehand Tennis serve Volleyball smash [Gupta et al, 2009] Slide Credit: Yao/Fei-Fei

49 Human Pose Estimation Results Method Torso Upper Leg Lower Leg Upper Arm Lower Arm Head Ramanan, 2006 Andriluka et al, 2009 Our full model Slide Credit: Yao/Fei-Fei

50 Human Pose Estimation Results Method Torso Upper Leg Lower Leg Upper Arm Lower Arm Head Ramanan, 2006 Andriluka et al, 2009 Our full model Tennis serve model Our estimation result Andriluka et al, 2009 Volleyball smash model Our estimation Andriluka result et al, 2009 Slide Credit: Yao/Fei-Fei

51 Human Pose Estimation Results Method Torso Upper Leg Lower Leg Upper Arm Lower Arm Head Ramanan, 2006 Andriluka et al, 2009 Our full model One pose per class Estimation result Estimation result Estimation result Estimation result Slide Credit: Yao/Fei-Fei

52 Dataset and Experiment Setup Sport data set: 6 classes 180 training & 120 testing images Tasks: Object detection; Pose estimation; Cricket defensive shot Cricket bowling Croquet shot Activity classification. Tennis forehand Tennis serve Volleyball smash [Gupta et al, 2009] Slide Credit: Yao/Fei-Fei

53 Activity Classification Results Classification accuracy % 78.9% 52.5% Cricket shot Tennis forehand 0.5 Our model Our model Gupta et al, 2009 Gupta et al, 2009 Bag-of- Words Bag-of-words SIFT+SVM Slide Credit: Yao/Fei-Fei

54 Motion features Dense Trajectory Action Recognition by Dense Trajectories, CVPR 2011 Action Recognition with Improved Trajectories, ICCV 2013

55 Video classification with CNNs Large-scale Video Classification with Convolutional Neural Networks, CVPR 2014

56 Video classification with CNNs Large-scale Video Classification with Convolutional Neural Networks, CVPR 2014

57 Two-stream CNN Two-Stream Convolutional Networks for Action Recognition in Videos, NIPS 2014

58 3D Convolutional Networks Learning Spatiotemporal Features with 3D Convolutional Networks, ICCV 2015

59 Action recognition -> Semantic role Labeling

60

61 Take-home messages Action recognition is an open problem. How to define actions? How to infer them? What are good visual cues? How do we incorporate higher level reasoning?

62 Take-home messages Some work done, but it is just the beginning of exploring the problem. So far Actions are mainly categorical (could be framed in terms of effect or intent) Most approaches are classification using simple features (spatial-temporal histograms of gradients or flow, s-t interest points, SIFT in images) Just a couple works on how to incorporate pose and objects Not much idea of how to reason about long-term activities or to describe video sequences