Building an Object Detection toolkit with TensorFlow (ODSC Europe 2017)

From academic papers to open source implementation.

In recent years, models based on Convolutional Neural Networks (CNNs) have revolutionized the entire field of computer vision. Problems like image classification can now be considered solved, and it is easy to construct implementations with any modern Deep Learning framework using fine tuning with pre-trained weights on datasets such as ImageNet. However, the harder problems of object detection and segmentation require much more complex methods to solve. Object detection consists of picking up the objects and drawing a rectangular bounding box, while segmentation aims to identify the exact pixels that belong to each object. One of the main differences with image classification is that the same image may contain several objects and those could be in very different proportions, sizes, lighting and partially occluded. In this talk, we will discuss how state of the art object detection techniques work. In particular, we will take a look at Region Proposal Networks (RPNs), which propose candidate object locations (“proposals”) which are later refined to achieve precise localization. We will then look at the architecture of an object detection system, and the performance considerations of different algorithms. Moreover, we will explore an implementation of an open source Python object detection toolkit based on TensorFlow, going through the details and tricks of the trade when using such ecosystem on production ready applications.

Transcript

1. @ODSC OPEN DATA SCIENCE CONFERENCE London | October 12th - 14th 2017
2. Building an Object Detection toolkit with TensorFlow From academic papers to open source implementation
3. | @tryolabs Who we are Javier Rey Lead Research Engineer @vierja Alan Descoins CTO @dekked_ 3 Introduction
4. Then vs now
5. 5 Introduction Felzenszwalb et. al., “Object Detection with Discriminatively Trained Part Based Models”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 32, 2010. Detected objects in a sample image (from the COCO dataset) (2017). Source: Google Research Blog. sofa bottle sofa
6. Agenda 6 Introduction 6 6 Introduction Challenges and applications of object detection Demystifying it: dive into Faster R-CNN Luminoth: our open-source toolkit for computer vision
7. Challenges of object detection 7 Introduction
8. Applications of object detection 8 Introduction CT scan of a lung cancer patient at the Jingdong Zhongmei private hospital in Yanjiao, China's Hebei Province (AP Photo/Andy Wong) Hsieh et. al., “Drone-based Object Counting by Spatially Regularized Regional Proposal Networks”, ICCV 2017. Source: Pinterest
9. A hard problem with lots of applications 9 Make it accessible! Build a toolkit! Introduction
10. Deep Learning & Object detection
11. Figure from https://blog.heuritech.com/2016/02/29/a-brief-report-of-the-heuritech-deep-learning-meetup-5/ Convolutional feature map Power of ConvNets as feature extractors Pre-train: 11 Deep Learning & Object detection
12. Regression based methods Type of object detection models 12 Deep Learning & Object detection Region proposal based methods Single stage prediction of object classes and bounding boxes. Examples: • You Only Look Once (YOLO) • Single Shot MultiBox Detector (SSD) Two stages: 1. Generate candidate locations using some algorithm. 2. Adjustment of bounding boxes and classification. Examples: • R-CNN, Fast R-CNN, Faster R-CNN
13. Faster R-CNN
14. Background Evolution of methods proposed in previous years: 14 Faster R-CNN 2014 R-CNN - Girshick et. al. 2015 Fast R-CNN - Girshick. 2016 Faster R-CNN Ren, Girshick et. al. “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks”, CVPR 2016.
15. 15 Faster R-CNN
16. 16 Overview Faster R-CNN RoIP 3. Region of Interest (RoI) Pooling R-CNN 4. Region-based CNN (R-CNN) RPN 2. Region Proposal Network (RPN) 1. Pre-trained base network
17. Faster R-CNN 1. Pre-trained base network
18. Pre-trained base network Image of arbitrary size → feature map. Common architectures: • VGG (16, 19) • ResNet (v1, v2) • Inception (V2, V3) Feature map encodes information for object detection. 18 Faster R-CNN Feature map 50 37 600 800 CNN (VGG16) 3 512
19. Faster R-CNN 2. Region Proposal Network (RPN)
20. Region proposals 20 Image (feature map) → proposals: • variable number. • different scales and aspect ratios. • efficient process • project bounding boxes to original image. Idea: start with reference boxes, later adjust. How many reference boxes? A lot! Faster R-CNN
21. Anchor boxes For each spatial position of the feature map, generate k fixed anchors (with same center). 21 Faster R-CNN 3 scales, 3 aspect ratios (k=9)
22. Anchor centers in original image Anchors reference (9 anchors per position) Visualizing anchor boxes (1) Anchors on top of single point 22 Faster R-CNN
23. All anchors superimposed Visualizing anchor boxes (2) Ground truth boxes labels: person, bicycle 23 Faster R-CNN
24. Region Proposal Network (RPN) Feature map → rectangular proposals + “objectness” score 24 Faster R-CNN RPN 3x3 conv (pad 1, 512 output channels) 1x1 conv (2k output channels) 1x1 conv (4k output channels) 2k objectness scores 4k box regression scores
25. All positive anchors IoU > 0.7 Anchors batch positive (green), negative (purple) 25 Faster R-CNN RPN anchor targets Need positive (foreground) vs negative (background) anchors. Use Intersection over Union (IoU) with ground truth. Faster R-CNN
26. What’s missing 26 Faster R-CNN Multi-task loss Filtering of proposals • Use Non-Maximum Suppression (NMS). • Keep top in “objectness” only. Classification Standard logarithmic loss for 2 classes. Box regression Smooth L1 between difference of coordinates (positive anchors).
27. 27 Faster R-CNN
28. Faster R-CNN 3. Region of Interest (RoI) Pooling
29. RoI Pooling layer 29 Faster R-CNN Arbitrarily-sized proposals → fixed spatial size • Can feed output to fully connected layers. • Very similar to max pooling. Faster R-CNN Project RoI Pool 7x7x512 Proposal RoI 512
30. Faster R-CNN 4. Region-based CNN (R-CNN)
31. Region-based CNN (R-CNN) 31 Faster R-CNN Fixed-size outputs of RoI Pooling→ Faster R-CNN 7x7x512 probability distribution (N+1 classes) bounding box regressions (N classes) Flatten FC FC bicycle p=0.96 Softmax
32. 32 Faster R-CNN person (0.99) bicycle (0.97)
33. Building a toolkit
34. Building a toolkit What is Luminoth? Open-source deep learning library/toolkit for computer vision object detection. 34 CLI tools Pre-defined models Cloud integration
35. $ pip install luminoth $ lumi train # Magic The goal 35 Building a toolkit
36. Building a toolkit Objectives 36 “Out-of-the-box” usage Production ready Open source Readable code Extensible and modular
37. Building a toolkit TensorFlow + Sonnet import sonnet as snt def RPN(snt.AbstractModule): def __init__(self, *args, name='rpn'): [...] # submodules init, config def _build(self, inputs): # TensorFlow code. return outputs 37 +
38. Building a toolkit “Model oriented programming” • Follow OOP good practices Faster R-CNN RPN in: → feature map, anchors out: → proposals RCNN in: → proposals, pooled feature maps out: → objects, labels, probabilities 38
39. ObjectDetection 39 Hierarchical structure Building a toolkit RPN R-CNN RoIP FasterRCNN RPN RPNTargets RPNProposals TFRecordDataset ObjectDetectionDataset RoIPooling RCNN RCNNTargets RCNNProposals TruncatedNetwork VGG/ResNet
40. Building a toolkit Challenges of coding from papers 40 Small implementation details have no room in academic papers Papers tend to remain frozen in time Many ways to implement it
41. Building a toolkit Challenges of Faster R-CNN implementation 41 Multiple moving parts Module dependencies Multi-task training
42. Data pipeline Debugging Training Data visualization Evaluation Deployment Beyond the model Distributed 42 Building a toolkit Unit testing Monitoring Model
43. Building a toolkit https://github.com/tryolabs/luminoth Using Luminoth 43 $ pip install luminoth $ lumi --help Usage: lumi [OPTIONS] COMMAND [ARGS]... Options: -h, --help Show this message and exit. Commands: cloud Groups of commands to train models in the cloud dataset Groups of commands to manage datasets evaluate Evaluate trained (or training) models server Groups of commands to serve models train Train models
44. $ lumi dataset transform --type pascal --data-dir /data/pascal --output /data/ # Create tfrecords for optimizing data consumption. $ lumi train --config pascal-fasterrcnn.yml # Hours of training $ tensorboard --logdir jobs/ # On another GPU/Machine/CPU $ lumi evaluate --config pascal-fasterrcnn.yml # Checks for new checkpoints and writes logs # Finally $ lumi server web --config pascal-fasterrcnn.yml # Looks for checkpoint and loads it into a simple frontend/json API server. Luminoth cycle 44 Building a toolkit
45. Building a toolkit Luminoth’s future 45 Fine-tune trained models More models & problems Tagging ↔ Training integration Distributed deployment
46. Thanks for listening! Questions? Learn more & contribute github.com/tryolabs/luminoth