Repository navigation
Expand file tree
/
Copy pathindex.html
More file actions
596 lines (556 loc) · 35.7 KB
/
Copy pathindex.html
File metadata and controls
596 lines (556 loc) · 35.7 KB
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
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
<!-- <!DOCTYPE html> -->
<html lang="en">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
<title>Yihui He, CMU grad student focused on Computer vision & Deep Learning</title>
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<!-- Le styles -->
<link href="./assets_files/bootstrap.min.css" rel="stylesheet">
<link href="./assets_files/bootstrap-responsive.min.css" rel="stylesheet">
<link href="./assets_files/yangqing.css" rel="stylesheet">
<!-- Le HTML5 shim, for IE6-8 support of HTML5 elements -->
<!--[if lt IE 9]>
<script src="assets/js/html5shiv.js"></script>
<![endif]-->
<link rel="icon" href="./assets_files/favicon.ico">
</head>
<div class="visible-phone" id="blackBar">
<a href="#top">About</a>
<!--<a href="#research">Research</a>-->
<a href="#publications">Publications</a>
<a href="#projects">Projects</a>
<!--<a href="#teaching">Teaching</a>-->
<a target="_blank"
href="https://drive.google.com/open?id=16ZSY62e2LgWYXuYuX4GVbbJ3OnU9bwl8">CV</a>
</div>
<body>
<div class="container span3 hidden-phone">
<div id="floating_sidebar" class="span3">
<!-- We use a fancy nav bar if there is enough space -->
<!--<hr class="hidden-phone">-->
<br>
<ul class="nav nav-list bs-docs-sidenav hidden-phone">
<li><a href="#top">About</a></li>
<!--<li><a href="#research">Research</a></li>-->
<li><a href="#publications">Publications</a></li>
<li><a href="#projects">Projects</a></li>
<!--<li><a href="#teaching">Teaching</a></li>-->
<li><a target="_blank"
href="https://drive.google.com/open?id=16ZSY62e2LgWYXuYuX4GVbbJ3OnU9bwl8">CV</a>
</li>
</ul>
<hr class="hidden-phone">
<div class="text-center hidden-phone">
<img src="assets_files/me.png" alt="photo" class="logo-image">
<br><br>
yihuihe.yh AT gmail.com <br>
</div>
<!-- Otherwise, we simply use a flat list of links -->
</div>
</div>
<div class="container">
<div class="row">
<div class="span9">
<br>
<h3>
Yihui He (何宜晖)
</h3>
<h5>
yihuihe.yh AT gmail DOT com
/ <a target="_blank"
href="https://scholar.google.com/citations?user=2yAMJ1YAAAAJ&hl=en">Google Scholar</a>
/ <a target="_blank"
href="https://github.com/yihui-he">GitHub</a>
/ <a target="_blank"
href="https://drive.google.com/open?id=16ZSY62e2LgWYXuYuX4GVbbJ3OnU9bwl8">CV</a>
/ <mark>actively looking for full-time / PhD position</mark>
</h5>
<!-- Do I want to show a pic on the phone screen?
<div class="text-center visible-phone">
<img src="assets/img/Yihui.png" alt="photo" width="150px"/>
</div>
-->
<a class="visible-phone pull-left" href="http://daggerfs.com/index.html#">
<img class="media-object" src="assets_files/me.png" width="96px" style="margin: 0px 10px">
</a>
<p>
I'm a CMU master student, with my interest focus on Computer Vision and Deep Learning.
Currently, I'm interning at Facebook AI Research mentored by <a target="_blank" href="https://research.fb.com/people/goyal-priya/">Priya Goyal</a> and <a target="_blank" href="http://www.cs.cmu.edu/~abhinavg/">Abhinav Gupta</a>.
At CMU, I focus on improving both <a href="#fsaf">one-stage</a>/<a href="#softer">two-stage</a> object detectors and hand pose estimation.
</p>
<p>
During my undergrad study I was fortunate to be an intern with <a target="_blank" href="http://jiansun.org/">Jian Sun</a> (Megvii/Face++), <a target="_blank" href="http://songhan.mit.edu">Song Han</a> (MIT) and <a target="_blank" href="http://www.cs.jhu.edu/~ayuille/">Alan Yuille</a> (JHU). I have a track record of contributing to CNN efficient inference.
Particularly, I designed <a href="#cp">channel pruning</a> to effectively prune channels.
I further proposed <a href="#AMC">AMC</a> to sample the design space of channel pruning via reinforcement learning, which greatly improved the performance.
</p>
<p>I served as a reviewer for ICCV'19, CVPR'19, ICLR'19, NIPS'18, TIP and IJCV. </p>
<!--
*** Research ***
-->
<!--<h3>-->
<!--<a name="research"></a> Research-->
<!--</h3>-->
<!--<p>-->
<!--My current research topics include:-->
<!--</p><ul>-->
<!--<li> Learning better structures for image feature extraction.-->
<!--</li><li> Explaining human generalization behavior with visually grounded cogscience models.-->
<!--</li><li> Making large-scale vision feasible and affordable.-->
<!--</li></ul>-->
<!--<p></p>-->
<!--<p> (Most recent publications to be added) </p>-->
<!--
*** Publications ***
-->
<h3>
<a name="publications"></a> Publications
</h3>
<div class="media">
<a name="pts" class="pull-left">
<img class="media-object" src="./assets/pts.png" width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>
Prediction-Tracking-Segmentation
</strong><br>
Jianren Wang, <strong>Yihui He</strong>, Xiaobo Wang, Xinjia Yu, Xia Chen
<a target="_blank"
href="https://arxiv.org/abs/1904.03280">[arXiv]</a>
</p>
<p class="abstract-text">
We motivate and present feature selective anchor-free(FSAF) module, a simple and effective building block forsingle-shot object detectors. It can be plugged into single-shot detectors with feature pyramid structure. The FSAF module addresses two limitations brought up by the conventional anchor-based detection: 1) heuristic-guided feature selection; 2) overlap-based anchor sampling.
</p>
</div>
</div>
<div class="media">
<a name="fsaf" class="pull-left">
<img class="media-object" src="https://ai2-s2-public.s3.amazonaws.com/figures/2017-08-08/791e7e83d86ef0f28a60ec6eb61dbc784d0300f4/7-Figure8-1.png" width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>
Feature Selective Anchor-Free Module for Single-Shot Object Detection
</strong><br>
Chenchen Zhu, <strong>Yihui He</strong>, <a target="_blank" href="http://www.cmu-biometrics.org/">Marios Savvides</a>, <strong>CVPR 2019</strong>
<a target="_blank"
href="https://arxiv.org/abs/1903.00621">[arXiv]</a>
</p>
<p class="abstract-text">
We motivate and present feature selective anchor-free (FSAF) module, a simple and effective building block forsingle-shot object detectors. It can be plugged into single-shot detectors with feature pyramid structure. The FSAF module addresses two limitations brought up by the conventional anchor-based detection: 1) heuristic-guided feature selection; 2) overlap-based anchor sampling.
</p>
</div>
</div>
<div class="media">
<a name="softer" class="pull-left">
<img class="media-object" src="https://raw.githubusercontent.com/yihui-he/softer-NMS/master/demo/output/softer.png" width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>
Bounding Box Regression with Uncertainty for Accurate Object Detection
</strong><br>
<strong>Yihui He</strong>, Chenchen Zhu, Jianren Wang, <a target="_blank" href="http://www.cmu-biometrics.org/">Marios Savvides</a>, Xiangyu Zhang, <strong>CVPR 2019</strong>
<a target="_blank"
href="https://arxiv.org/abs/1809.08545">[arXiv]</a>
<a target="_blank"
href="https://www.youtube.com/watch?v=bcGtNdTzdkc">[presentation]</a>
<a target="_blank"
href="https://github.com/yihui-he/KL-Loss">[code]</a>
</p>
<p class="abstract-text">
We introduce a novel bounding box regression loss for learning bounding box transformation and localization variance together. The resulting localization variance is utilized in our new non-maximum suppression method to improve localization accuracy for object detection. On MS-COCO, we boost the AP of VGG-16 faster R-CNN from 23.6% to 29.1% with a single model and nearly no additional computational overhead. More importantly, our method improves the AP of ResNet-50 FPN fast R-CNN from 36.8% to 37.8%, which achieves state-of-the-art bounding box refinement result.
</p>
</div>
</div>
<div class="media">
<a name="address" class="pull-left">
<img class="media-object" src="https://raw.githubusercontent.com/yihui-he/yihui-he.github.io/master/assets/Screen%20Shot%202018-10-03%20at%2001.53.35.png" width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>
AddressNet: Shift-based Primitives for Efficient Convolutional Neural Networks
</strong><br>
<strong>Yihui He*</strong>, Xianggen Liu*, Huasong Zhong* and Yuchun Ma, <strong>WACV 2019</strong>
<a target="_blank"
href="https://arxiv.org/pdf/1809.08458.pdf">[arXiv]</a>
<a target="_blank"
href="https://www.youtube.com/watch?v=2Abi_razAXI">[presentation]</a>
<a target="_blank"
href="https://raw.githubusercontent.com/yihui-he/yihui-he.github.io/master/assets/WACV19-poster.png">[poster]</a>
<a target="_blank"
href="https://github.com/yihui-he/yihui-he.github.io/blob/master/assets/AddressNet_%20Shift-based%20Primitives%20for%20Efficient%20CNN.pptx?raw=true">[slides]</a>
</p>
<p class="abstract-text">
We propose a collection of three shift-based primitives for building efficient compact CNN-based networks. These three primitives (channel shift, address shift, shortcut shift) can reduce the inference time on GPU while maintains the prediction accuracy. These shift-based primitives only moves the pointer but avoids memory copy, thus very fast.
</p>
</div>
</div>
<div class="media">
<a name="AMC" class="pull-left">
<img class="media-object" src="./assets_files/mobile.png" width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>
AMC: AutoML for Model Compression and Acceleration on Mobile Devices
</strong><br>
<strong>Yihui He*</strong>, Ji Lin*, Zhijian Liu, Hanrui Wang, Li-Jia Li, <a target="_blank" href="http://songhan.mit.edu">Song Han</a>, <strong>ECCV 2018</strong>
<a target="_blank"
href="http://openaccess.thecvf.com/content_ECCV_2018/html/Yihui_He_AMC_Automated_Model_ECCV_2018_paper.html">[PDF]</a>
<a target="_blank"
href="https://arxiv.org/abs/1802.03494">[arXiv]</a>
<a target="_blank"
href="https://github.com/mit-han-lab/amc-compressed-models">[code]</a>
</p>
<p class="abstract-text">
In this paper, we propose AMC: AutoML for Model Compression that leverage reinforcement learning to efficiently sample the design space and greatly improve the model compression quality.
</p>
</div>
</div>
<div class="media">
<a name="cp" class="pull-left">
<img class="media-object" src="./assets_files/ill-1.png" width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>
Channel pruning for accelerating very deep neural networks
</strong><br>
<strong>Yihui He</strong>, Xiangyu Zhang, <a target="_blank" href="http://jiansun.org/">Jian Sun</a>, <strong>ICCV 2017</strong>
<a target="_blank"
href="http://openaccess.thecvf.com/content_iccv_2017/html/He_Channel_Pruning_for_ICCV_2017_paper.html">[PDF]</a>
<a target="_blank"
href="https://arxiv.org/abs/1707.06168">[arXiv]</a>
<a target="_blank"
href="https://raw.githubusercontent.com/yihui-he/images/master/38722_He_0600.png">[poster]</a>
<a target="_blank"
href="http://nbviewer.jupyter.org/github/yihui-he/images/blob/master/channel-pruning-methods.pdf">[slides]</a>
<a target="_blank"
href="https://github.com/yihui-he/channel-pruning">[code]</a>
</p>
<p class="abstract-text">
In this paper, we introduce a new channel pruning method to accelerate very deep convolutional neural networks.Given a trained CNN model, we propose an iterative two-step algorithm to effectively prune each layer, by a LASSO regression based channel selection and least square reconstruction. We further generalize this algorithm to multi-layer and multi-branch cases. Our method reduces the accumulated error and enhance the compatibility with various architectures. Our pruned VGG-16 achieves the state-of-the-art results by 5x speed-up along with only 0.3% increase of error. More importantly, our method is able to accelerate modern networks like ResNet, Xception and suffers only 1.4%, 1.0% accuracy loss under 2x speed-up respectively, which is significant.
</p>
</div>
</div>
<div class="media">
<a class="pull-left">
<img class="media-object" src="./assets_files/vehicle.png" width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>Vehicle Traffic Driven Camera Placement for Better Metropolis Security
Surveillance</strong><br>
Xiaobo Ma*, <strong>Yihui He*</strong>, Xiapu Luo, Jianfeng Li, Xiaohong Guan
, <strong>IEEE Intelligent System</strong>
<a target="_blank"
href="https://ieeexplore.ieee.org/abstract/document/8354911/">[PDF]</a>
<a target="_blank"
href="http://arxiv.org/abs/1705.08508">[arXiv]</a>
<a target="_blank"
href="https://github.com/yihui-he/Vehicle-Traffic-Driven-Camera-Placement">[code]</a>
</p>
<p class="abstract-text">
Security surveillance is one of the most important issues in smart cities, especially in an era
of
terrorism. Deploying a number of (video) cameras is a
common approach for surveillance information retrieval.
Given the never-ending power offered by vehicles to a
metropolis, exploiting vehicle traffic to design camera
placement strategies could potentially facilitate physicalworld security surveillance. We take
the first step towards
exploring the linkage between vehicle traffic and camera
placement in favor of physical-world security surveillance
from a network perspective.
</p>
</div>
</div>
<div class="media">
<a class="pull-left">
<img class="media-object" src="./assets_files/filterwise.png" width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>
Pruning Very Deep Neural Network Channels for Efficient Inference
</strong><br>
<strong>Yihui He</strong>, Xiangyu Zhang, <a target="_blank" href="http://jiansun.org/">Jian Sun</a>
, <i>TPAMI, Major Revision</i>
</p>
<p class="abstract-text">
Channel Pruning is further expanded to Filterwise Pruning with rich experiements.
</p>
</div>
</div>
<div class="media">
<a class="pull-left">
<img class="media-object" src="./assets_files/superres.png" width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>
Single Image Super-resolution with a Parameter Economic Residual-like Convolutional Neural Network
</strong><br>
Yudong Liang, Ze Yang, Kai Zhang, <strong>Yihui He</strong>, Jinjun Wang, Nanning Zheng
<a target="_blank"
href="https://arxiv.org/abs/1703.08173">[arXiv]</a>
</p>
<p class="abstract-text">
This paper aims to extend the merits of residual network, such as skip
connection induced fast training, for a typical low-level vision problem,
i.e., single image super-resolution. In general, the two main challenges
of existing deep CNN for supper-resolution lie in the gradient exploding/vanishing
problem and large amount of parameters or computational
cost as CNN goes deeper. Correspondingly, the skip connections or identity
mapping shortcuts are utilized to avoid gradient exploding/vanishing
problem. To tackle with the second problem, a parameter economic CNN
architecture which has carefully designed width, depth and skip connections
was proposed. Different residual-like architectures for image superresolution
has also been compared. Experimental results have demonstrated
that the proposed CNN model can not only achieve state-of-the-art
PSNR and SSIM results for single image super-resolution but also produce
visually pleasant results.
</p>
</div>
</div>
<!--<div class="media">-->
<!--<a class="pull-left" href="#top">-->
<!--<img class="media-object" src="./assets_files/decaf-features.png" width="96px" height="96px">-->
<!--</a>-->
<!--<div class="media-body">-->
<!--<p class="media-heading">-->
<!--<strong>DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition</strong><br>-->
<!--J Donahue, Y Jia, O Vinyals, J Hoffman, N Zhang, E Tzeng, T Darrell. arXiv preprint.<br>-->
<!--<a target="_blank" href="http://arxiv.org/abs/1310.1531">[ArXiv Link]</a>-->
<!--<a target="_blank" href="http://decaf.berkeleyvision.org/">[Live Demo]</a>-->
<!--<a target="_blank" href="https://github.com/UCB-ICSI-Vision-Group/decaf-release/">[Software]</a>-->
<!--<a target="_blank" href="http://www.eecs.berkeley.edu/~jiayq/decaf_pretrained/">[Pretrained ImageNet Model]</a>-->
<!--</p>-->
<!--<p class="abstract-text">-->
<!--We evaluate whether features extracted from the activation of a deep convolutional network trained in a fully supervised fashion on a large, fixed set of object recognition tasks can be re-purposed to novel generic tasks. We also released the software and pre-trained network to do large-scale image classification.-->
<!--</p>-->
<!--</div>-->
<!--</div>-->
<!--
*** Projects ***
-->
<h3>
<a name="projects"></a> Projects
</h3>
Link to my <a target="_blank" href="https://github.com/yihui-he">[github public projects]</a>
<div class="media">
<a class="pull-left">
<img class="media-object"
src="https://raw.githubusercontent.com/yihui-he/TSP-report/master/02132.png"
width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>
An Empirical Analysis of Approximation Algorithms for the Euclidean Traveling Salesman Problem
</strong>
<a target="_blank"
href="https://arxiv.org/pdf/1705.09058.pdf">[PDF]</a>
<a target="_blank"
href="https://github.com/yihui-he/TSP">[Code]</a>
</p>
<p class="abstract-text">
we perform an evaluation and analysis of cornerstone algorithms for the metric TSP. We evaluate greedy, 2-opt, and genetic algorithms. We use
several datasets as input for the algorithms including a small dataset, a medium-sized dataset representing
cities in the United States, and a synthetic dataset consisting of 200 cities to test algorithm scalability.
</p>
</div>
</div>
<div class="media">
<a class="pull-left">
<img class="media-object"
src="https://raw.githubusercontent.com/yihui-he/Depth-estimation-with-neural-network/master/presentation/stereo.png?token=AJkBS_A-YWaMd9vcgEQuaXQWe9wmjtTBks5XWM07wA%3D%3D"
width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>
Estimated Depth Map Helps Image Classification
</strong>
<a target="_blank"
href="https://arxiv.org/pdf/1709.07077.pdf">[PDF]</a>
<a target="_blank" href="https://github.com/yihui-he/Estimated-Depth-Map-Helps-Image-Classification">[Code]</a>
<a target="_blank"
href="http://nbviewer.jupyter.org/github/yihui-he/Estimated-Depth-Map-Helps-Image-Classification/blob/master/presentation/depth.pdf">[Slides]</a>
</p>
<p class="abstract-text">
We consider image classification with estimated depth. This problem falls into the domain of transfer learning, since we are using a model trained on a set of depth images to generate depth maps (additional features) for use in another classification problem using another disjoint set of images. It's challenging as no direct depth information is provided. Though depth estimation has been well studied, none have attempted to aid image classification with estimated depth. Therefore, we present a way of transferring domain knowledge on depth estimation to a separate image classification task over a disjoint set of train, and test data. We build a RGBD dataset based on RGB dataset and do image classification on it. Then evaluation the performance of neural networks on the RGBD dataset compared to the RGB dataset. From our experiments, the benefit is significant with shallow and deep networks. It improves ResNet-20 by 0.55% and ResNet-56 by 0.53%.
</p>
</div>
</div>
<div class="media">
<a class="pull-left">
<img class="media-object"
src="https://www.kaggle.io/svf/310043/4567203286d71c8fd31cf12668a3ceac/__results___files/__results___5_0.png"
width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>
Medical Image Segmentation: A survey
</strong>
<a target="_blank" href="https://github.com/yihui-he/u-net">[Code]</a>
<a target="_blank"
href="https://github.com/yihui-he/medical-image-segmentation-a-survey/blob/master/README.md">[Summary]</a>
</p>
<p class="abstract-text">
I evaluate DeepMask, Deeplab and MNC for medical image segmentation. I ranked
<strong>19%</strong> on <a target="_blank"
href="https://www.kaggle.com/c/ultrasound-nerve-segmentation/leaderboard/private">Kaggle
Ultrasound Nerve Segmentation</a>
</p>
</div>
</div>
<div class="media">
<a class="pull-left">
<img class="media-object"
src="https://raw.githubusercontent.com/yihui-he/deep-learning-guide/master/presentation/resnet.png"
width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>
residual neural network with tensorflow on CIFAR-100
</strong>
<a target="_blank"
href="http://nbviewer.jupyter.org/github/yihui-he/ResNet-tensorflow/blob/master/report/mp2_Yihui%20He.pdf">[PDF]</a>
<a target="_blank" href="https://github.com/yihui-he/ResNet-tensorflow">[Code]</a>
<a target="_blank"
href="http://nbviewer.jupyter.org/github/yihui-he/deep-learning-guide/blob/master/presentation/mp12.pdf">[Slides]</a>
</p>
<p class="abstract-text">
I reimplement 6/13/20 layers RNN in tensorflow from scratch, and test it’s result on
CIFAR10/100.
</p>
</div>
</div>
<div class="media">
<a class="pull-left">
<img class="media-object"
src="https://raw.githubusercontent.com/yihui-he/deep-learning-guide/master/presentation/kmeans.jpg"
width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>Single Layer neural network with PCAwhitening Kmeans</strong>
<a target="_blank" href="http://nbviewer.jupyter.org/github/yihui-he/An-Analysis-of-Single-Layer-Networks-in-Unsupervised-Feature-Learning/blob/master/report/mp1_Yihui%20He.pdf">[PDF]</a>
<a target="_blank"
href="https://github.com/yihui-he/An-Analysis-of-Single-Layer-Networks-in-Unsupervised-Feature-Learning">[Code]</a>
<a target="_blank"
href="http://nbviewer.jupyter.org/github/yihui-he/deep-learning-guide/blob/master/presentation/mp12.pdf">[Slides]</a>
</p>
<p class="abstract-text">
We evaluate whether features extracted from the activation of a deep convolutional network
trained in a fully supervised fashion on a large, fixed set of object recognition tasks can be
re-purposed to novel generic tasks. We also released the software and pre-trained network to do
large-scale image classification.
</p>
</div>
</div>
<div class="media">
<a class="pull-left">
<img class="media-object"
src="https://raw.githubusercontent.com/yihui-he/my-Notes/master/person.jpg.png" width="96px"
height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>Objects Detection with YOLO on Artwork Dataset</strong>
<a target="_blank" href="http://nbviewer.jupyter.org/github/yihui-he/Objects-Detection-with-YOLO-on-Artwork-Dataset/blob/master/Report_Yihui.pdf">[PDF]</a>
<a target="_blank"
href="https://github.com/yihui-he/Objects-Detection-with-YOLO-on-Artwork-Dataset">[Code]</a>
</p>
<p class="abstract-text">
I design a small object detection network, which is simplified from YOLO(You Only Look Once)
network. It's trained on PASCAL VOC. I evaluate it on an artwork dataset(Picasso dataset). With
the best parameters, I got 40% precision and 35% recall.
</p>
</div>
</div>
<div class="media">
<a class="pull-left">
<img class="media-object" src="./assets_files/shuttle.png" width="96px" height="96px">
</a>
<div class="media-body">
<p class="media-heading">
<strong>shuttlecock detection and tracking</strong>
<a target="_blank"
href="https://github.com/yihui-he/Badminton-Robot">[Code]</a>
</p>
<p class="abstract-text">
With guassian mixture model, I extract shuttlecock proposals. Then I use Partical filter to
refine proposals. From multi view cameras, I employed
structure from motion to predict its 3D location. Combined with Physics laws, landing location
prediction accuracy is around 5 cm. (This system
works on embeded linux with openCV)
</p>
</div>
</div>
<!-- Footer
================================================== -->
<hr>
<footer class="footer">
<div class='hidden-phone'>
<!-- <h3 class="text-center"><a name="wall"></a><strong>works</strong></h3>-->
<section id="photos">
<img src="https://raw.githubusercontent.com/yihui-he/lip-tracking-with-snake-active-contour-and-particle-filter/master/pic.png"/>
<img src="https://raw.githubusercontent.com/yihui-he/Edge-detection-with-zero-crossing/master/lena_1.bmp"/>
<img src="https://raw.githubusercontent.com/yihui-he/3D-reconstruction/master/result/selfff.png"/>
<img src="./assets_files/cs188.png"/>
<img src="https://raw.githubusercontent.com/yihui-he/my-Notes/master/person.jpg.png"/>
<img src="./assets_files/vehicle.png"/>
<img src="https://raw.githubusercontent.com/yihui-he/Depth-estimation-with-neural-network/master/presentation/stereo.png?token=AJkBS_A-YWaMd9vcgEQuaXQWe9wmjtTBks5XWM07wA%3D%3D"/>
<img src="https://raw.githubusercontent.com/yihui-he/deep-learning-guide/master/presentation/resnet.png"/>
<img src="https://raw.githubusercontent.com/yihui-he/deep-learning-guide/master/presentation/kmeans.jpg"/>
<img src="./assets_files/shuttle.png"/>
<img src="https://raw.githubusercontent.com/yihui-he/my-Notes/master/gaoxin.jpg"/>
<img src="https://raw.githubusercontent.com/yihui-he/my-Notes/master/artwork.jpg"/>
<img src="https://raw.githubusercontent.com/yihui-he/my-Notes/master/bop.jpg"/>
<img src="./assets_files/ocsi.png"/>
</section>
<a target="_blank" href="https://github.com/yihui-he/panorama"><img
src="https://github.com/yihui-he/panorama/blob/master/results/yellowstone5.jpg?raw=true"></a>
<hr>
</div>
<div class="row">
<div class="span12">
<p>
modified from <a target="_blank" href="http://daggerfs.com/">© Yangqing Jia 2013</a>
</p>
</div>
</div>
</footer>
</div>
</div>
</div>
</body>
</html>
<!-- Le javascript
================================================== -->
<!-- Placed at the end of the document so the pages load faster -->
<!--
<script type="text/javascript" src="http://platform.twitter.com/widgets.js"></script>
<script src="//ajax.googleapis.com/ajax/libs/jquery/1.9.1/jquery.min.js"></script>
<script src="assets/js/bootstrap-transition.js"></script>
<script src="assets/js/bootstrap-alert.js"></script>
<script src="assets/js/bootstrap-modal.js"></script>
<script src="assets/js/bootstrap-dropdown.js"></script>
<script src="assets/js/bootstrap-scrollspy.js"></script>
<script src="assets/js/bootstrap-tab.js"></script>
<script src="assets/js/bootstrap-tooltip.js"></script>
<script src="assets/js/bootstrap-popover.js"></script>
<script src="assets/js/bootstrap-button.js"></script>
<script src="assets/js/bootstrap-collapse.js"></script>
<script src="assets/js/bootstrap-carousel.js"></script>
<script src="assets/js/bootstrap-typeahead.js"></script>
<script src="assets/js/bootstrap-affix.js"></script>
<script src="assets/js/holder/holder.js"></script>
<script src="assets/js/application.js"></script>
-->