-
Notifications
You must be signed in to change notification settings - Fork 26
/
Copy path2_object_detection_slides.html
389 lines (242 loc) · 11.3 KB
/
2_object_detection_slides.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
<!DOCTYPE html>
<html>
<head>
<title>Object detection</title>
<meta charset="utf-8">
<meta name="author" content="Sigrid Keydana" />
<link rel="stylesheet" href="theme/rstudio.css" type="text/css" />
</head>
<body>
<textarea id="source">
class: center, middle, inverse, title-slide
# Object detection
### Sigrid Keydana
### rstudio::conf 2019
---
# Learning objectives
- Perform classification and localization on a single object
- Important concepts in multiple-object detection
- Understand how to code a _very_ basic version of SSD (__Single-Shot Multibox Detector__)
- Understand the options to improve on this basic detector
---
class: inverse, middle, center
# Road to object detection
---
# Single-object classification and localization
- Mix of recap (if you've participated in yesterday's workshop) and new topics
- Partly demo, partly exercise
- We'll also look at the dataset and preprocessing required for this session
---
# PASCAL Visual Object Classes (VOC) challenges and datasets
- Challenges (2005-2012) included
- classification (presence/absence for each object class)
- object detection (same as above, plus localization)
- class segmentation
- "person layout"
- We'll use the training set from the [2007 challenge](http://host.robots.ox.ac.uk/pascal/VOC/voc2007/index.html)
- Number of object classes: 20
- Number of training images: 2501
- We focus on concepts and the how-to, not accuracy
---
# Object detection examples
---
# How do you learn bounding boxes?
- Can be framed as __regression problem__
- often trained with mean absolute error
- Predict pixel coordinates of box corners (_x_left_, _y_top_, _x_right_, _y_bottom_)
- Relevant metric is __Intersection over Union__ (__IOU__), also known as Jaccard index<sup>1</sup>
.footnote[[1] Image source: Wikipedia.]
---
# Demo/exercise: Single-object classification and localization
- Notebook: [2_object_detection/1_classification_localization.Rmd](2_object_detection/1_classification_localization.Rmd)
- Quiz: [2_object_detection/object_detection_quizzes.Rmd](2_object_detection/object_detection_quizzes.Rmd)
---
class: inverse, middle, center
# Introduction to multiple-object detection
---
# Why can't we just have more bounding boxes?
Each bounding box detector will try to detect all objects:<sup>1</sup>
We need to either:
- zoom in on single objects in some way, or
- have detectors __specialize__ in what to detect
.footnote[[1] Image source: http://machinethink.net/blog/object-detection/]
---
# Object detection main approaches / paradigms
- Sliding windows approaches
- Train network, run sequentially on image patches
- May actually run sliding windows synchronously (see _Overfeat_ below)
- Region proposal approaches (2-step)
- Step 1: Some algorithm proposes interesting regions
- Step 2: Another algorithm (a convnet) classifies the regions and refines localization
- Single-shot detectors (YOLO, SSD)
- Perform detection, classification and localization in one step
---
# Sliding windows done synchronously
.footnote[[1] cf. Sermanet, P, D. Eigen, X. Zhang, et al. (2013). OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks. ]
---
# Region proposal approaches
- R-CNN<sup>1</sup>: Uses non-DL algorithm to select interesting regions, then applies CNN to all identified regions sequentially
- Fast R-CNN<sup>2</sup>: Uses non-DL algorithm to select interesting regions, then classifies all regions in one pass
- Faster R-CNN<sup>3</sup>: Uses a convnet for region proposal (_Region proposal network_), then classifies all regions in one pass
<span class="footnote">
[1] cf. Girshick, R. B. (2015). Fast R-CNN.<br />
[2] cf. Girshick, R. B, J. Donahue, T. Darrell, et al. (2013). Rich feature hierarchies for accurate object detection and semantic segmentation.<br />
[3] cf. Ren, S, K. He, R. B. Girshick, et al. (2015). Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks.
</span>
---
# Single-shot detectors
Make detectors __specialize__ using __anchor boxes__<sup>1</sup>
.footnote[[1] illustration from: Liu, W, D. Anguelov, D. Erhan, et al. (2015). SSD: Single Shot MultiBox Detector.]
---
# Why do anchor boxes help?
### Concept 1: Grid
By laying a grid over the image, we _map detectors to image regions_.
### Concept 2: Different aspect ratios and scales
Objects of different shapes are localized more easily because we lean on different customized _priors_.
<br />
<img src="2_object_detection/images/ssd_1.png" width = "50%">
---
# SSD architecture
<br />
How do we detect objects that are of completely different sizes than a grid cell (e.g., span the whole image)?
SSD adds anchor boxes at different resolutions:
<img src="2_object_detection/images/ssd_arch.png" width = "120%">
---
# SSD vs. YOLO
- Early YOLO versions (v1<sup>1</sup>/v2<sup>2</sup>) had dense layers before the final output, but current YOLO<sup>3</sup> is fully convolutional just like SSD
- In addition to _number of classes_ class scores and 4 bounding box coordinates __per detector__, YOLO also has a _confidence score_
- YOLO determines anchor boxes for each dataset individually (using k-means clustering on the actual bounding boxes), while SSD uses _default boxes_
- Details of bounding box computations differ
<span class="footnote">
[1] cf. Redmon, J, S. K. Divvala, R. B. Girshick, et al. (2015). You Only Look Once: Unified, Real-Time Object Detection. <br />
[2] cf. Redmon, J. and A. Farhadi (2016). YOLO9000: Better, Faster, Stronger. <br />
[3] cf. Redmon, J. and A. Farhadi (2018). YOLOv3: An Incremental Improvement.
</span>
---
# Why fully convolutional?
As long as for every detector, the part of the image it's responsible for is visible (= is in its receptive field), we don't need a fully connected layer at the end<sup>1</sup>
<br />
<img src="2_object_detection/images/receptive_field.png" width = "50%"/>
.footnote[[1] Image from: Goodfellow, I, Y. Bengio and A. Courville (2016). Deep Learning.]
---
class: inverse, middle, center
# Coding a (very!) basic single-shot detector
---
# Basic SSD: Code
- To show the basic approach, we will
- restrict ourselves to a 4x4 grid of image cells
- have one anchor box per cell (thus, 16 anchor boxes)
- don't aggregate detections from different resolutions
- Notebook: [2_object_detection/2_object_detection_ssd.Rmd](2_object_detection/2_object_detection_ssd.Rmd)
---
# Basic SSD: Ways for improvement
- Use focal loss
- Use anchor boxes of different aspect ratios
- Perform detection at various resolutions
- (not object detection specific:) Use data augmentation
---
# Focal Loss<sup>1</sup>
<img src="2_object_detection/images/focal.png" width = "70%"/>
.footnote[[1] Image from: Lin, T, P. Goyal, R. B. Girshick, et al. (2017). Focal Loss for Dense Object Detection.]
---
# Focal loss, and more anchors
<br />
Implementation stubs:
- Focal loss: [2_object_detection/3_object_detection_ssd_focal.Rmd](2_object_detection/3_object_detection_ssd_focal.Rmd)
- More anchors: [2_object_detection/4_object_detection_ssd_moreanchors.Rmd](2_object_detection/4_object_detection_ssd_moreanchors.Rmd)
---
# Wrapup / feedback
---
# References
Girshick, R. B. (2015). "Fast R-CNN". In: _CoRR_ abs/1504.08083.
eprint: 1504.08083. URL:
[http://arxiv.org/abs/1504.08083](http://arxiv.org/abs/1504.08083).
Girshick, R. B, J. Donahue, T. Darrell, et al. (2013). "Rich
feature hierarchies for accurate object detection and semantic
segmentation". In: _CoRR_ abs/1311.2524. eprint: 1311.2524. URL:
[http://arxiv.org/abs/1311.2524](http://arxiv.org/abs/1311.2524).
Goodfellow, I, Y. Bengio and A. Courville (2016). _Deep Learning_.
<URL: http://www.deeplearningbook.org>. MIT Press.
Lin, T, P. Goyal, R. B. Girshick, et al. (2017). "Focal Loss for
Dense Object Detection". In: _CoRR_ abs/1708.02002. eprint:
1708.02002. URL:
[http://arxiv.org/abs/1708.02002](http://arxiv.org/abs/1708.02002).
Liu, W, D. Anguelov, D. Erhan, et al. (2015). "SSD: Single Shot
MultiBox Detector". In: _CoRR_ abs/1512.02325. eprint: 1512.02325.
URL:
[http://arxiv.org/abs/1512.02325](http://arxiv.org/abs/1512.02325).
---
# References (cont.)
Redmon, J, S. K. Divvala, R. B. Girshick, et al. (2015). "You Only
Look Once: Unified, Real-Time Object Detection". In: _CoRR_
abs/1506.02640. eprint: 1506.02640. URL:
[http://arxiv.org/abs/1506.02640](http://arxiv.org/abs/1506.02640).
Redmon, J. and A. Farhadi (2016). "YOLO9000: Better, Faster,
Stronger". In: _CoRR_ abs/1612.08242. eprint: 1612.08242. URL:
[http://arxiv.org/abs/1612.08242](http://arxiv.org/abs/1612.08242).
Redmon, J. and A. Farhadi (2018). "YOLOv3: An Incremental
Improvement". In: _CoRR_ abs/1804.02767. eprint: 1804.02767. URL:
[http://arxiv.org/abs/1804.02767](http://arxiv.org/abs/1804.02767).
Ren, S, K. He, R. B. Girshick, et al. (2015). "Faster R-CNN:
Towards Real-Time Object Detection with Region Proposal Networks".
In: _CoRR_ abs/1506.01497. eprint: 1506.01497. URL:
[http://arxiv.org/abs/1506.01497](http://arxiv.org/abs/1506.01497).
Sermanet, P, D. Eigen, X. Zhang, et al. (2013). "OverFeat:
Integrated Recognition, Localization and Detection using
Convolutional Networks". In: _CoRR_ abs/1312.6229. eprint:
1312.6229. URL:
[http://arxiv.org/abs/1312.6229](http://arxiv.org/abs/1312.6229).
</textarea>
<script src="https://remarkjs.com/downloads/remark-latest.min.js"></script>
<script>var slideshow = remark.create({
"highlightStyle": "github",
"highlightLines": true,
"countIncrementalSlides": false,
"ratio": "16:9"
});
if (window.HTMLWidgets) slideshow.on('afterShowSlide', function (slide) {
window.dispatchEvent(new Event('resize'));
});
(function() {
var d = document, s = d.createElement("style"), r = d.querySelector(".remark-slide-scaler");
if (!r) return;
s.type = "text/css"; s.innerHTML = "@page {size: " + r.style.width + " " + r.style.height +"; }";
d.head.appendChild(s);
})();</script>
<script>
(function() {
var i, text, code, codes = document.getElementsByTagName('code');
for (i = 0; i < codes.length;) {
code = codes[i];
if (code.parentNode.tagName !== 'PRE' && code.childElementCount === 0) {
text = code.textContent;
if (/^\\\((.|\s)+\\\)$/.test(text) || /^\\\[(.|\s)+\\\]$/.test(text) ||
/^\$\$(.|\s)+\$\$$/.test(text) ||
/^\\begin\{([^}]+)\}(.|\s)+\\end\{[^}]+\}$/.test(text)) {
code.outerHTML = code.innerHTML; // remove <code></code>
continue;
}
}
i++;
}
})();
</script>
<!-- dynamically load mathjax for compatibility with self-contained -->
<script>
(function () {
var script = document.createElement('script');
script.type = 'text/javascript';
script.src = 'https://mathjax.rstudio.com/latest/MathJax.js?config=TeX-MML-AM_CHTML';
if (location.protocol !== 'file:' && /^https?:/.test(script.src))
script.src = script.src.replace(/^https?:/, '');
document.getElementsByTagName('head')[0].appendChild(script);
})();
</script>
</body>
</html>