mirror of
https://github.com/docling-project/docling-core.git
synced 2026-11-18 20:27:00 +00:00
* New labels and utility for CVAT annotation Signed-off-by: Christoph Auer <cau@zurich.ibm.com> * Add EMPTY_VALUE to TextItem Signed-off-by: Christoph Auer <cau@zurich.ibm.com> * Small fixes Signed-off-by: Christoph Auer <cau@zurich.ibm.com> * align usage of content layer param (#326) * align usage of content layer param Signed-off-by: Panos Vagenas <pva@zurich.ibm.com> * parametrize content layers in visualizers Signed-off-by: Panos Vagenas <pva@zurich.ibm.com> --------- Signed-off-by: Panos Vagenas <pva@zurich.ibm.com> --------- Signed-off-by: Christoph Auer <cau@zurich.ibm.com> Signed-off-by: Panos Vagenas <pva@zurich.ibm.com> Signed-off-by: Christoph Auer <60343111+cau-git@users.noreply.github.com> Co-authored-by: Panos Vagenas <35837085+vagenas@users.noreply.github.com>
541 lines
28 KiB
Plaintext
Vendored
541 lines
28 KiB
Plaintext
Vendored
0: unspecified with name=_root_
|
|
1: page_header: arXiv:2206.01062v1 [cs.CV] 2 Jun 2022
|
|
2: section_header: DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis
|
|
3: text: Birgit Pfitzmann IBM Research Rueschlikon, Switzerland bpf@zurich.ibm.com
|
|
4: text: Christoph Auer IBM Research Rueschlikon, Switzerland cau@zurich.ibm.com
|
|
5: text: Michele Dolfi IBM Research Rueschlikon, Switzerland dol@zurich.ibm.com
|
|
6: text: Ahmed S. Nassar IBM Research Rueschlikon, Switzerland ahn@zurich.ibm.com
|
|
7: text: Peter Staar IBM Research Rueschlikon, Switzerland taa@zurich.ibm.com
|
|
8: section_header: ABSTRACT
|
|
9: text: Accurate document layout analysis is a key requirement for highquality PDF document conversion. With
|
|
10: section_header: CCS CONCEPTS
|
|
11: text: · Informationsystems → Documentstructure ; · Appliedcomputing → Document analysis ; · Computing meth
|
|
12: text: Permission to make digital or hard copies of part or all of this work for personal or classroom use
|
|
13: text: KDD '22, August 14-18, 2022, Washington, DC, USA
|
|
14: text: © 2022 Copyright held by the owner/author(s).
|
|
15: text: ACM ISBN 978-1-4503-9385-0/22/08.
|
|
16: text: https://doi.org/10.1145/3534678.3539043
|
|
17: picture
|
|
18: caption: Figure 1: Four examples of complex page layouts across different document categories
|
|
19: text: 13
|
|
20: text: USING THE VERTICAL TUBE -
|
|
21: text: MODELS AY11230/11234
|
|
22: text: 1. The vertical tube can be used for
|
|
23: text: ional viewing or to photograph
|
|
24: text: instruct
|
|
25: text: the image with a digital camera or a
|
|
26: text: micro TV unit
|
|
27: text: 2. Loosen the retention screw, then rotate
|
|
28: text: the adjustment ring to change the
|
|
29: text: re that both the images in
|
|
30: text: length of the vertical tube.
|
|
31: text: 3. Make su
|
|
32: text: OPERATION (cont.)
|
|
33: text: SELECTING OBJECTIVE
|
|
34: text: MAGNIFICATION
|
|
35: text: 1. There are two objectives. The lower
|
|
36: text: magnification objective has a greater
|
|
37: text: depth of field and view.
|
|
38: text: 2. In order to observe the specimen
|
|
39: text: easily use the lower magnification
|
|
40: text: objective first. Then, by rotating the
|
|
41: text: case, the magnification can be
|
|
42: text: changed.
|
|
43: text: CHANGING THE INTERPUPILLARY
|
|
44: text: DISTANCE
|
|
45: text: 1. The distance between the observer's
|
|
46: text: pupils is the interpupil ary distance.
|
|
47: text: 2. To adjust the interpupil ary distance
|
|
48: text: rotate the prism caps until both eyes
|
|
49: text: coincide with the image in the
|
|
50: text: eyepiece.
|
|
51: text: FOCUSING
|
|
52: text: 1. Remove the lens protective cover.
|
|
53: text: 2. Place the specimen on the working
|
|
54: text: stage.
|
|
55: text: 3. Focus the specimen with the left eye
|
|
56: text: first while turning the focus knob until
|
|
57: text: the image appears clear and sharp.
|
|
58: text: 4. Rotate the right eyepiece ring until the
|
|
59: text: images in each eyepiece coincide and
|
|
60: text: are sharp and clear.
|
|
61: text: CHANGING THE BULB
|
|
62: text: 1. Disconnect the power cord.
|
|
63: text: 2. When the bulb is cool, remove the
|
|
64: text: oblique il uminator cap and remove
|
|
65: text: the halogen bulb with cap.
|
|
66: text: 3. Replace with a new halogen bulb.
|
|
67: text: 4. Open the window in the base plate and
|
|
68: text: replace the halogen lamp or
|
|
69: text: fluorescent lamp of transmitted
|
|
70: text: il uminator.
|
|
71: text: FOCUSING
|
|
72: text: 1. Turn the focusing knob away or toward
|
|
73: text: you until a clear image is viewed.
|
|
74: text: 2. If the image is unclear, adjust the
|
|
75: text: height of the elevator up or down,
|
|
76: text: then turn the focusing knob again.
|
|
77: text: ZOOM MAGNIFICATION
|
|
78: text: 1. Turn the zoom magnification knob to
|
|
79: text: the desired magnification and field of
|
|
80: text: view.
|
|
81: text: 2. In most situations, it is recommended
|
|
82: text: that you focus at the lowest
|
|
83: text: magnification, then move to a higher
|
|
84: text: magnification and re-focus as
|
|
85: text: necessary.
|
|
86: text: 3. If the image is not clear to both eyes
|
|
87: text: at the same time, the diopter ring may
|
|
88: text: need adjustment.
|
|
89: text: DIOPTER RING ADJUSTMENT
|
|
90: text: 1. To adjust the eyepiece for viewing with
|
|
91: text: or without eyeglasses and for
|
|
92: text: differences in acuity between the right
|
|
93: text: and left eyes, fol ow the fol owing
|
|
94: text: steps:
|
|
95: text: a. Observe an image through the left
|
|
96: text: eyepiece and bring a specific point
|
|
97: text: into focus using the focus knob.
|
|
98: text: b. By turning the diopter ring
|
|
99: text: adjustment for the left eyepiece,
|
|
100: text: bring the same point into sharp
|
|
101: text: focus.
|
|
102: text: c.Then bring the same point into
|
|
103: text: focus through the right eyepiece
|
|
104: text: by turning the right diopter ring.
|
|
105: text: d.With more than one viewer, each
|
|
106: text: viewer should note their own
|
|
107: text: diopter ring position for the left
|
|
108: text: and right eyepieces, then before
|
|
109: text: viewing set the diopter ring
|
|
110: text: adjustments to that setting.
|
|
111: text: CHANGING THE BULB
|
|
112: text: 1. Disconnect the power cord from the
|
|
113: text: electrical outlet.
|
|
114: text: 2. When the bulb is cool, remove the
|
|
115: text: oblique il uminator cap and remove
|
|
116: text: the halogen bulb with cap.
|
|
117: text: 3. Replace with a new halogen bulb.
|
|
118: text: 4. Open the window in the base plate
|
|
119: text: and replace the halogen lamp or
|
|
120: text: fluorescent lamp of transmitted
|
|
121: text: il uminator.
|
|
122: text: Model AY11230
|
|
123: text: Model AY11234
|
|
124: text: 14
|
|
125: text: Objectives
|
|
126: text: Revolving Turret
|
|
127: text: Coarse
|
|
128: text: Adjustment
|
|
129: text: Knob
|
|
130: text: MODEL AY11236
|
|
131: text: MICROSCOPE USAGE
|
|
132: text: BARSKA Model AY11236 is a powerful fixed power compound
|
|
133: text: microscope designed for biological studies such as specimen
|
|
134: text: examination. It can also be used for examining bacteria and
|
|
135: text: for general clinical and medical studies and other scientific uses.
|
|
136: text: CONSTRUCTION
|
|
137: text: BARSKA Model AY11236 is a fixed power compound microscope.
|
|
138: text: It is constructed with two optical paths at the same angle. It is
|
|
139: text: equipped with transmitted il umination. By using this instrument,
|
|
140: text: the user can observe specimens at magnification from 40x to
|
|
141: text: 1000x by selecting the desired objective lens. Coarse and fine
|
|
142: text: focus adjustments provide accuracy and image detail. The rotating
|
|
143: text: head al ows the user to position the eyepieces for maximum
|
|
144: text: viewing comfort and easy access to al adjustment knobs.
|
|
145: text: Model AY11236
|
|
146: text: Fine
|
|
147: text: Adjustment
|
|
148: text: Knob
|
|
149: text: Stage
|
|
150: text: Condenser
|
|
151: text: Focusing
|
|
152: text: Knob
|
|
153: text: Eyepiece
|
|
154: text: Stand
|
|
155: text: Lamp
|
|
156: text: On/Of
|
|
157: text: Switch
|
|
158: text: Lamp
|
|
159: text: Power
|
|
160: text: Cord
|
|
161: text: Rotating Head
|
|
162: text: Stage Clip
|
|
163: text: Adjustment
|
|
164: text: Interpupil ary Slide Adjustment
|
|
165: text: Circling Minimums
|
|
166: text: 7KHUHglyph<c=3,font=/AAAAAH+ArialMT>ZDVglyph<c=3,font=/AAAAAH+ArialMT>Dglyph<c=3,font=/AAAAAH+ArialM
|
|
167: text: improved obstacle protection. To indicate that the new criteria had been applied to a given procedur
|
|
168: text: is placed on
|
|
169: text: the circling line of minimums. The new circling tables and explanatory information is located in the
|
|
170: text: 7KHglyph<c=3,font=/AAAAAH+ArialMT>DSSURDFKHVglyph<c=3,font=/AAAAAH+ArialMT>XVLQJglyph<c=3,font=/AAAA
|
|
171: text: on the circling line of
|
|
172: text: minima.
|
|
173: text: $SSO\glyph<c=3,font=/AAAAAK+Arial-ItalicMT>6WDQGDUGglyph<c=3,font=/AAAAAK+Arial-ItalicMT>&LUFOLQJgly
|
|
174: text: $SSO\glyph<c=3,font=/AAAAAK+Arial-ItalicMT>([SDQGHGglyph<c=3,font=/AAAAAK+Arial-ItalicMT>&LUFOLQJgly
|
|
175: text: Table
|
|
176: text: AIRPORT SKETCH
|
|
177: text: The airport sketch is a depiction of the airport with emphasis on runway pattern and related
|
|
178: text: information, positioned in either the lower left or lower right corner of the chart to aid pi-
|
|
179: text: lot recognition of the airport from the air and to provide some information to aid on ground
|
|
180: text: navigation of the airport. The runways are drawn to scale and oriented to true north. Runway
|
|
181: text: dimensions (length and width) are shown for al active runways.
|
|
182: text: Runway(s) are depicted based on what type and construction of the runway.
|
|
183: text: Hard Surface
|
|
184: text: Other Than
|
|
185: text: Hard Surface
|
|
186: text: Metal Surface
|
|
187: text: Closed Runway
|
|
188: text: Under Construction
|
|
189: text: Stopways,
|
|
190: text: Taxiways, Park-
|
|
191: text: ing Areas
|
|
192: text: Displaced
|
|
193: text: Threshold
|
|
194: text: Closed
|
|
195: text: Pavement
|
|
196: text: Water Runway
|
|
197: text: Taxiways and aprons are shaded grey. Other runway features that may be shown are runway numbers, run
|
|
198: text: sions, runway slope, arresting gear, and displaced threshold.
|
|
199: text: 2WKHUglyph<c=3,font=/AAAAAH+ArialMT>LQIRUPDWLRQglyph<c=3,font=/AAAAAH+ArialMT>FRQFHUQLQJglyph<c=3,fo
|
|
200: text: -
|
|
201: text: pads may also be shown.
|
|
202: text: $LUSRUWglyph<c=3,font=/AAAAAN+Arial-BoldMT>(OHYDWLRQglyph<c=3,font=/AAAAAN+Arial-BoldMT>DQGglyph<c=3
|
|
203: text: The airport elevation is shown enclosed within a box in the upper left corner of the sketch box and
|
|
204: text: 114
|
|
205: text: FAA Chart Users' Guide - Terminal Procedures Publication (TPP) - Terms
|
|
206: text: AGL 2013 Financial Calendar
|
|
207: text: 22 August 2012
|
|
208: text: 2012 ful year result and final dividend announced
|
|
209: text: 30 August 2012
|
|
210: text: Ex-dividend trading commences
|
|
211: text: 5 September 2012
|
|
212: text: Record date for 2012 final dividend
|
|
213: text: 27 September 2012
|
|
214: text: Final dividend payable
|
|
215: text: 23 October 2012
|
|
216: text: Annual General Meeting
|
|
217: text: 27 February 2013 1
|
|
218: text: 2013 interim result and interim dividend announced
|
|
219: text: 28 August 2013 1
|
|
220: text: 2013 ful year results and final dividend announced
|
|
221: text: 1 Indicative dates only, subject to change/Board confirmation
|
|
222: text: AGL's Annual General Meeting wil be held at the City Recital Hal , Angel Place, Sydney
|
|
223: text: commencing at 10.30am on Tuesday 23 October 2012.
|
|
224: text: Yesterday
|
|
225: text: Established in Sydney in 1837, and then
|
|
226: text: known as The Australian Gas Light Company,
|
|
227: text: the AGL business has an established history
|
|
228: text: and reputation for serving the gas and
|
|
229: text: electricity needs of Australian households.
|
|
230: text: In 1841, when AGL supplied the gas to light
|
|
231: text: the first public street lamp, it was reported
|
|
232: text: in the Sydney Gazette as a 'wonderful
|
|
233: text: achievement of scientific knowledge, assisted
|
|
234: text: by mechanical ingenuity.' Within two years,
|
|
235: text: 165 gas lamps were lighting the City of Sydney.
|
|
236: text: Looking back on
|
|
237: text: 175years of
|
|
238: text: lookingforward.
|
|
239: text: AGL Energy Limited ABN 74 115 061 375
|
|
240: text: 29
|
|
241: text: signs, signals and road markings
|
|
242: text: 3
|
|
243: text: In
|
|
244: text: chapter 2, you and your vehicle
|
|
245: text: , you learned about
|
|
246: text: some of the controls in your vehicle. This chapter is a handy
|
|
247: text: reference section that gives examples of the most common
|
|
248: text: signs, signals and road markings that keep traffic organized
|
|
249: text: and flowing smoothly.
|
|
250: text: Signs
|
|
251: text: There are three ways to read signs: by their shape, colour and
|
|
252: text: the messages printed on them. Understanding these three ways
|
|
253: text: of classifying signs wil help you figure out the meaning of signs
|
|
254: text: that are new to you.
|
|
255: text: Stop
|
|
256: text: Yield the right-of-way
|
|
257: text: Shows driving
|
|
258: text: regulations
|
|
259: text: Explains lane use
|
|
260: text: School zone signs
|
|
261: text: are fluorescent
|
|
262: text: yel ow-green
|
|
263: text: Tel s about motorist
|
|
264: text: services
|
|
265: text: Shows a permitted
|
|
266: text: action
|
|
267: text: Shows an action that
|
|
268: text: is not permitted
|
|
269: text: Warns of hazards
|
|
270: text: ahead
|
|
271: text: Warns of
|
|
272: text: construction zones
|
|
273: text: Railway crossing
|
|
274: text: Shows distance and
|
|
275: text: direction
|
|
276: text: • Signs
|
|
277: text: -
|
|
278: text: regulatory signs
|
|
279: text: -
|
|
280: text: school,
|
|
281: text: playground and
|
|
282: text: crosswalk signs
|
|
283: text: -
|
|
284: text: lane use signs
|
|
285: text: -
|
|
286: text: urn control signs
|
|
287: text: t
|
|
288: text: -
|
|
289: text: parking signs
|
|
290: text: -
|
|
291: text: eserved lane
|
|
292: text: r
|
|
293: text: signs
|
|
294: text: -
|
|
295: text: warning signs
|
|
296: text: -
|
|
297: text: object markers
|
|
298: text: -
|
|
299: text: onstruction
|
|
300: text: c
|
|
301: text: signs
|
|
302: text: -
|
|
303: text: information and
|
|
304: text: destination signs
|
|
305: text: -
|
|
306: text: ailway signs
|
|
307: text: r
|
|
308: text: • Signals
|
|
309: text: -
|
|
310: text: lane control
|
|
311: text: signals
|
|
312: text: -
|
|
313: text: traffic lights
|
|
314: text: • Road markings
|
|
315: text: -
|
|
316: text: yel ow lines
|
|
317: text: -
|
|
318: text: white lines
|
|
319: text: -
|
|
320: text: reserved lane
|
|
321: text: markings
|
|
322: text: -
|
|
323: text: other markings
|
|
324: text: in this chapter
|
|
325: section_header: KEYWORDS
|
|
326: text: PDF document conversion, layout segmentation, object-detection, data set, Machine Learning
|
|
327: section_header: ACMReference Format:
|
|
328: text: Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. 2022. DocLayNet:
|
|
329: page_header: KDD '22, August 14-18, 2022, Washington, DC, USA Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ah
|
|
330: section_header: 1 INTRODUCTION
|
|
331: text: Despite the substantial improvements achieved with machine-learning (ML) approaches and deep neural
|
|
332: text: Akeyproblem in the process of document conversion is to understand the structure of a single documen
|
|
333: text: In this paper, we present the DocLayNet dataset. It provides pageby-page layout annotation ground-tr
|
|
334: list with name=list
|
|
335: list_item: (1) Human Annotation : In contrast to PubLayNet and DocBank, we relied on human annotation instead o
|
|
336: list_item: (2) Large Layout Variability : We include diverse and complex layouts from a large variety of public
|
|
337: list_item: (3) Detailed Label Set : We define 11 class labels to distinguish layout features in high detail. Pu
|
|
338: list_item: (4) Redundant Annotations : A fraction of the pages in the DocLayNet data set carry more than one hu
|
|
339: footnote: 1 https://developer.ibm.com/exchanges/data/all/doclaynet
|
|
340: text: This enables experimentation with annotation uncertainty and quality control analysis.
|
|
341: list with name=list
|
|
342: list_item: (5) Pre-defined Train-, Test- & Validation-set : Like DocBank, we provide fixed train-, test- & vali
|
|
343: text: All aspects outlined above are detailed in Section 3. In Section 4, we will elaborate on how we desi
|
|
344: text: In Section 5, we will present baseline accuracy numbers for a variety of object detection methods (F
|
|
345: section_header: 2 RELATED WORK
|
|
346: text: While early approaches in document-layout analysis used rulebased algorithms and heuristics [8], the
|
|
347: text: Lately, new types of ML models for document-layout analysis have emerged in the community [18-21]. T
|
|
348: section_header: 3 THE DOCLAYNET DATASET
|
|
349: text: DocLayNet contains 80863 PDF pages. Among these, 7059 carry two instances of human annotations, and
|
|
350: text: In addition to open intellectual property constraints for the source documents, we required that the
|
|
351: page_header: DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis
|
|
352: page_header: KDD '22, August 14-18, 2022, Washington, DC, USA
|
|
353: picture
|
|
354: caption: Figure 2: Distribution of DocLayNet pages across document categories.
|
|
355: text: Patents
|
|
356: text: 8%
|
|
357: text: Scientific
|
|
358: text: 17%
|
|
359: text: Financial
|
|
360: text: 32%
|
|
361: text: Tenders
|
|
362: text: 6%
|
|
363: text: Laws
|
|
364: text: 16%
|
|
365: text: Manuals
|
|
366: text: 21%
|
|
367: text: to a minimum, since they introduce difficulties in annotation (see Section 4). As a second condition
|
|
368: text: The pages in DocLayNet can be grouped into six distinct categories, namely Financial Reports , Manua
|
|
369: text: We did not control the document selection with regard to language. The vast majority of documents co
|
|
370: text: To ensure that future benchmarks in the document-layout analysis community can be easily compared, w
|
|
371: footnote: 2 e.g. AAPL from https://www.annualreports.com/
|
|
372: text: Table 1 shows the overall frequency and distribution of the labels among the different sets. Importa
|
|
373: text: In order to accommodate the different types of models currently in use by the community, we provide
|
|
374: text: Despite being cost-intense and far less scalable than automation, human annotation has several benef
|
|
375: section_header: 4 ANNOTATION CAMPAIGN
|
|
376: text: The annotation campaign was carried out in four phases. In phase one, we identified and prepared the
|
|
377: page_header: KDD '22, August 14-18, 2022, Washington, DC, USA Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ah
|
|
378: table
|
|
379: caption: Table 1: DocLayNet dataset overview. Along with the frequency of each class label, we present the re
|
|
380: picture
|
|
381: caption: Figure 3: Corpus Conversion Service annotation user interface. The PDF page is shown in the backgrou
|
|
382: text: Text
|
|
383: text: Picture
|
|
384: text: Formula
|
|
385: text: Code
|
|
386: text: Complex-form
|
|
387: text: Section-header
|
|
388: text: footer
|
|
389: text: Page-header
|
|
390: text: Footnote
|
|
391: text: Table
|
|
392: text: Caption
|
|
393: text: List-item
|
|
394: text: Title
|
|
395: text: Skip
|
|
396: text: Filter
|
|
397: text: Submit
|
|
398: text: Report problem
|
|
399: text: Page
|
|
400: text: we distributed the annotation workload and performed continuous quality controls. Phase one and two
|
|
401: text: Phase 1: Data selection and preparation. Our inclusion criteria for documents were described in Sect
|
|
402: text: Preparation work included uploading and parsing the sourced PDF documents in the Corpus Conversion S
|
|
403: text: Phase 2: Label selection and guideline. We reviewed the collected documents and identified the most
|
|
404: footnote: 3 https://arxiv.org/
|
|
405: page_header: DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis
|
|
406: page_header: KDD '22, August 14-18, 2022, Washington, DC, USA
|
|
407: text: the textual content of an element, which goes beyond visual layout recognition, in particular outsid
|
|
408: text: At first sight, the task of visual document-layout interpretation appears intuitive enough to obtain
|
|
409: text: Obviously, this inconsistency in annotations is not desirable for datasets which are intended to be
|
|
410: list with name=list
|
|
411: list_item: (1) Every list-item is an individual object instance with class label List-item . This definition is
|
|
412: list_item: (2) A List-item is a paragraph with hanging indentation. Singleline elements can qualify as List-ite
|
|
413: list_item: (3) For every Caption , there must be exactly one corresponding Picture or Table .
|
|
414: list_item: (4) Connected sub-pictures are grouped together in one Picture object.
|
|
415: list_item: (5) Formula numbers are included in a Formula object.
|
|
416: list_item: (6) Emphasised text (e.g. in italic or bold) at the beginning of a paragraph is not considered a Sec
|
|
417: text: The complete annotation guideline is over 100 pages long and a detailed description is obviously out
|
|
418: text: Phase 3: Training. After a first trial with a small group of people, we realised that providing the
|
|
419: picture
|
|
420: text: 1ef23f5e6d7f10d393f9947e8208285dce9ae87250ac483ac4b4a59d51b4e037
|
|
421: text: Compliant with guidelines
|
|
422: text: Plausible but invalid alternative
|
|
423: text: Borderline case: Two guideline-compliant alternatives
|
|
424: text: 03c31a2ee1ed1b583c28957f475ee545d144e1b5a264dc4dd068c8d2f6a64860
|
|
425: text: 1a5cd524f1844c1260c8e8c073e1f442423c264583212b0d0b6626fc780e6ed4
|
|
426: text: A
|
|
427: text: B
|
|
428: text: C
|
|
429: text: D
|
|
430: text: 05237a14f2524e3f53c8454b074409d05078038a6a36b770fcc8ec7e540deae0
|
|
431: caption: Figure 4: Examples of plausible annotation alternatives for the same page. Criteria in our annotatio
|
|
432: text: were carried out over a timeframe of 12 weeks, after which 8 of the 40 initially allocated annotator
|
|
433: text: Phase 4: Production annotation. The previously selected 80K pages were annotated with the defined 11
|
|
434: page_header: KDD '22, August 14-18, 2022, Washington, DC, USA Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ah
|
|
435: text: Table 2: Prediction performance (mAP@0.5-0.95) of object detection networks on DocLayNet test set. T
|
|
436: table
|
|
437: text: to avoid this at any cost in order to have clear, unbiased baseline numbers for human document-layou
|
|
438: section_header: 5 EXPERIMENTS
|
|
439: text: The primary goal of DocLayNet is to obtain high-quality ML models capable of accurate document-layou
|
|
440: picture
|
|
441: caption: Figure 5: Prediction performance (mAP@0.5-0.95) of a Mask R-CNNnetworkwithResNet50backbonetrainedoni
|
|
442: text: 0
|
|
443: text: 20
|
|
444: text: 40
|
|
445: text: 60
|
|
446: text: 80
|
|
447: text: 100
|
|
448: text: % of DocLayNet training set
|
|
449: text: 50
|
|
450: text: 55
|
|
451: text: 60
|
|
452: text: 65
|
|
453: text: 70
|
|
454: text: mAP 0.50:0.95
|
|
455: text: 10 1
|
|
456: text: 10 2
|
|
457: text: 50
|
|
458: text: 55
|
|
459: text: 60
|
|
460: text: 65
|
|
461: text: 70
|
|
462: text: paper and leave the detailed evaluation of more recent methods mentioned in Section 2 for future wor
|
|
463: text: In this section, we will present several aspects related to the performance of object detection mode
|
|
464: section_header: Baselines for Object Detection
|
|
465: text: In Table 2, we present baseline experiments (given in mAP) on Mask R-CNN [12], Faster R-CNN [11], an
|
|
466: page_header: DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis
|
|
467: page_header: KDD '22, August 14-18, 2022, Washington, DC, USA
|
|
468: text: Table 3: Performance of a Mask R-CNN R50 network in mAP@0.5-0.95 scores trained on DocLayNet with di
|
|
469: table
|
|
470: section_header: Learning Curve
|
|
471: text: One of the fundamental questions related to any dataset is if it is 'large enough'. To answer this q
|
|
472: section_header: Impact of Class Labels
|
|
473: text: The choice and number of labels can have a significant effect on the overall model performance. Sinc
|
|
474: text: Table 4: Performance of a Mask R-CNN R50 network with document-wise and page-wise split for differen
|
|
475: table
|
|
476: text: lists in PubLayNet (grouped list-items) versus DocLayNet (separate list-items), the label set of siz
|
|
477: section_header: Impact of Document Split in Train and Test Set
|
|
478: text: Many documents in DocLayNet have a unique styling. In order to avoid overfitting on a particular sty
|
|
479: section_header: Dataset Comparison
|
|
480: text: Throughout this paper, we claim that DocLayNet's wider variety of document layouts leads to more rob
|
|
481: page_header: KDD '22, August 14-18, 2022, Washington, DC, USA Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ah
|
|
482: text: Table 5: Prediction Performance (mAP@0.5-0.95) of a Mask R-CNN R50 network across the PubLayNet, Doc
|
|
483: table
|
|
484: text: Section-header , Table and Text . Before training, we either mapped or excluded DocLayNet's other la
|
|
485: text: For comparison of DocBank with DocLayNet, we trained only on Picture and Table clusters of each data
|
|
486: section_header: Example Predictions
|
|
487: text: To conclude this section, we illustrate the quality of layout predictions one can expect from DocLay
|
|
488: section_header: 6 CONCLUSION
|
|
489: text: In this paper, we presented the DocLayNet dataset. It provides the document conversion and layout an
|
|
490: text: From the dataset, we have derived on the one hand reference metrics for human performance on documen
|
|
491: text: To date, there is still a significant gap between human and ML accuracy on the layout interpretation
|
|
492: section_header: REFERENCES
|
|
493: list with name=list
|
|
494: list_item: [1] Max Göbel, Tamir Hassan, Ermelinda Oro, and Giorgio Orsi. Icdar 2013 table competition. In 2013
|
|
495: list_item: [2] Christian Clausner, Apostolos Antonacopoulos, and Stefan Pletschacher. Icdar2017 competition on
|
|
496: list_item: [3] Hervé Déjean, Jean-Luc Meunier, Liangcai Gao, Yilun Huang, Yu Fang, Florian Kleber, and Eva-Mari
|
|
497: list_item: [4] Antonio Jimeno Yepes, Peter Zhong, and Douglas Burdick. Competition on scientific literature par
|
|
498: list_item: [5] Logan Markewich, Hao Zhang, Yubin Xing, Navid Lambert-Shirzad, Jiang Zhexin, Roy Lee, Zhi Li, an
|
|
499: list_item: [6] Xu Zhong, Jianbin Tang, and Antonio Jimeno-Yepes. Publaynet: Largest dataset ever for document l
|
|
500: list_item: [7] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A b
|
|
501: list_item: [8] Riaz Ahmad, Muhammad Tanvir Afzal, and M. Qadir. Information extraction from pdf sources based o
|
|
502: list_item: [9] Ross B. Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for
|
|
503: list_item: [10] Ross B. Girshick. Fast R-CNN. In 2015 IEEE International Conference on Computer Vision , ICCV,
|
|
504: list_item: [11] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object d
|
|
505: list_item: [12] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross B. Girshick. Mask R-CNN. In IEEE Internati
|
|
506: list_item: [13] Glenn Jocher, Alex Stoken, Ayush Chaurasia, Jirka Borovec, NanoCode012, TaoXie, Yonghye Kwon, K
|
|
507: page_header: DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis
|
|
508: page_header: KDD '22, August 14-18, 2022, Washington, DC, USA
|
|
509: picture
|
|
510: caption: Text Caption List-Item Formula Table Section-Header Picture Page-Header Page-Footer Title
|
|
511: text: 4bed2a8aa51ac37058e79605821bbc426d032b0b6ca8bdf3409ed8508ccd8c67
|
|
512: text: MOTION
|
|
513: text: 2f2a06d08f5ad565d0f5e815f4ddf666365b2cff435cdaeb8850217e8a8efabf
|
|
514: text: W
|
|
515: text: 7f2fd7293e04bf4f1756ae51f5779764933da1d1d2002e3915356050570fc75b
|
|
516: text: OPERATION (cont.
|
|
517: text: 1b81cf65f47456ad4faa725d1eb09879bd633af16cfe2bf8cea661b87907bfac
|
|
518: text: b60da9d26f488cb133e47d101d35fda1bdca2671ade60764d1cd569590270327
|
|
519: text: 2b7b8355a42ebef0cf91583aad9f30f7c9fa63c5b05911730ba15275c024965b
|
|
520: text: A
|
|
521: text: B
|
|
522: text: C
|
|
523: text: D
|
|
524: text: E
|
|
525: text: moti
|
|
526: text: F
|
|
527: text: Figure 6: Example layout predictions on selected pages from the DocLayNet test-set. (A, D) exhibit f
|
|
528: text: Diaconu, Mai Thanh Minh, Marc, albinxavi, fatih, oleg, and wanghao yang. ultralytics/yolov5: v6.0 -
|
|
529: list with name=list
|
|
530: list_item: [20] Shoubin Li, Xuyan Ma, Shuaiqun Pan, Jun Hu, Lin Shi, and Qing Wang. Vtlayout: Fusion of visual
|
|
531: list_item: [14] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Ser
|
|
532: list_item: [15] Mingxing Tan, Ruoming Pang, and Quoc V. Le. Efficientdet: Scalable and efficient object detecti
|
|
533: list_item: [16] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hay
|
|
534: list_item: [17] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2, 2019.
|
|
535: list_item: [18] Nikolaos Livathinos, Cesar Berrospi, Maksym Lysak, Viktor Kuropiatnyk, Ahmed Nassar, Andre Carv
|
|
536: list_item: [19] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutlm: Pre-training
|
|
537: list_item: [21] Peng Zhang, Can Li, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, and Fei Wu. Vsr: A unified
|
|
538: list_item: [22] Peter W J Staar, Michele Dolfi, Christoph Auer, and Costas Bekas. Corpus conversion service: A
|
|
539: list_item: [23] Connor Shorten and Taghi M. Khoshgoftaar. A survey on image data augmentation for deep learning
|