Correction: Icariin and icaritin recover UVB-induced photoaging by stimulating Nrf2/ARE and reducing AP-1 and NF-κB signaling pathways: a comparative study on UVB-irradiated human keratinocytes.

Correction for 'Icariin and icaritin recover UVB-induced photoaging by stimulating Nrf2/ARE and reducing AP-1 and NF-κB signaling pathways: a comparative study on UVB-irradiated human keratinocytes' by Eunson Hwang et al., Photochem. Photobiol. Sci., 2018, 17, 1396-1408.

Icariin and icaritin recover UVB-induced photoaging by stimulating Nrf2/ARE and reducing AP-1 and NF-κB signaling pathways: a comparative study on UVB-irradiated human keratinocytes.

Icariin (ICA) and icaritin (ICT) exhibit many pharmacological functions including anti-osteoporosis, anti-cardiovascular, and anti-cancer activities; however, there are few comprehensive studies that track the detailed effects on UVB-induced photoaging. The recovery effects of ICA and ICT were investigated in UVB-irradiated human keratinocytes (HaCaTs). The results indicated that ICT and ICA showed strong radical scavenging activity, and the reactive oxygen species (ROS) scavenging activity of ICT was superior. UVB-induced matrix metalloproteinase-1 (MMP-1) expression was blocked by ICA via the inhibition of mitogen-activated protein kinase/activator protein 1 (MAPK/AP-1), which directly reduced extracellular matrix (ECM) degradation. ICT activated nuclear factor erythroid 2 related factor 2 (Nrf2) to improve the anti-oxidative stress capacity and suppress nuclear factor-κB (NF-κB) activation, decreasing vascular endothelial growth factor (VEGF) protein, and inflammatory cytokines induced ECM degrading enzyme secretion. Moreover, ICT was more advantageous to improve transforming growth factor beta 1 (TGF-ß1) and procollagen type I expression than ICA, promoting the synthesis of collagen. Therefore, ICA and ICT have potential to treat UVB-induced oxidative stress, inflammation and photoaging, and will be posited as a novel strategy to alleviate photodamage.

Conversion of Ginsenoside Rb1 into Six Types of Highly Bioactive Ginsenoside Rg3 and Its Derivatives by FeCl3 Catalysis.

Ginsenoside Rb1 is an important saponin of ginseng(s); however, Rb1, with 3-O- and 20-O-sugar moieties, has low bioavailability. Here, we report the derivatization of ginsenoside Rb1 to completely generate six types of highly bioactive minor ginsenoside Rg3 and its derivatives by FeCl3 catalysis, the reaction conditions are similar to enzymatic reaction conditions. In FeCl3 catalysis, the only 20-O-sugar-moiety of ginsenoside Rb1 was decomposed into the minor ginsenosides Rk1 and Rg5 with newly produced C-20 ethylene bands; but also hydrolyzed into 20(S)-Rg3 and 20(R)-Rg3; subsequently the C-24(25) ethylene bands of 20(S)-Rg3 and 20(R)-Rg3 were hydrated to 20(S)-25-OH-Rg3 and 20(R)-25-OH-Rg3. After separation of reaction mixture from 34 g ginsenoside-Rb1 by silica-gel-column, the 3.3 g sample I of TLC top-band consisting of Rg5 and Rk1, 8.7 g sample II of TLC middle-band consisting of 20(S)-Rg3 and 20(R)-Rg3, 3.5 g sample III of TLC bottom-band consisting of unknown product-I and -II including 20(S)-25-OH-Rg3, were obtained. The sample III consisting of unknown product-I and -II was purified by crystallization, and identified to 20(S)-25-OH-Rg3 and 20(R)-25-OH-Rg3 by HPLC-Evaporative Light Scattering Detector (ELSD) and NMR. Therefore, six types of minor-ginsenosides Rk1, Rg5, 20(S)-Rg3, 20(R)-Rg3, 20(S)-25-OH-Rg3 and 20(R)-25-OH-Rg3 were successfully prepared from ginsenoside Rb1 by FeCl3 catalysis. FeCl3 has low toxicity and is inexpensive, and the reaction conditions are similar to enzymatic reaction conditions; thus, this method is applicable to the development of ginseng-based drugs.

Automatic Camera Calibration Using Active Displays of a Virtual Pattern.

Camera calibration plays a critical role in 3D computer vision tasks. The most commonly used calibration method utilizes a planar checkerboard and can be done nearly fully automatically. However, it requires the user to move either the camera or the checkerboard during the capture step. This manual operation is time consuming and makes the calibration results unstable. In order to solve the above problems caused by manual operation, this paper presents a full-automatic camera calibration method using a virtual pattern instead of a physical one. The virtual pattern is actively transformed and displayed on a screen so that the control points of the pattern can be uniformly observed in the camera view. The proposed method estimates the camera parameters from point correspondences between 2D image points and the virtual pattern. The camera and the screen are fixed during the whole process; therefore, the proposed method does not require any manual operations. Performance of the proposed method is evaluated through experiments on both synthetic and real data. Experimental results show that the proposed method can achieve stable results and its accuracy is comparable to the standard method by Zhang.

Acinetobacter plantarum sp. nov. isolated from wheat seedlings plant.

Strain THG-SQM11(T), a Gram-negative, aerobic, non-motile, coccus-shaped bacterium, was isolated from wheat seedlings plant in P. R. China. Strain THG-SQM11(T) was closely related to members of the genus Acinetobacter and showed the highest 16S rRNA sequence similarities with Acinetobacter junii (97.9 %) and Acinetobacter kookii (96.1 %). DNA-DNA hybridization showed 41.3 ± 2.4 % DNA reassociation with A. junii KCTC 12416(T). Chemotaxonomic data revealed that strain THG-SQM11(T) possesses ubiquinone-9 as the predominant respiratory quinone, C18:1 ω9c, summed feature 3 (C16:1 ω7c and/or C16:1 ω6c), and C16:0 as the major fatty acids. The major polar lipids were found to be diphosphatidylglycerol, phosphatidylethanolamine, phosphatidylglycerol, and phosphatidylcholine. The DNA G+C content was 41.7 mol %. These data, together with phenotypic characterization, suggest that the isolate represents a novel species, for which the name Acinetobacter plantarum sp. nov. is proposed, with THG-SQM11(T) as the type strain (=CCTCC AB 2015123(T) =KCTC 42611(T)).

Mucilaginibacter pocheonensis sp. nov., with ginsenoside-converting activity, isolated from soil of a ginseng-cultivating field.

A Gram-reaction-negative, aerobic, heterotrophic, non-motile, non-spore-forming, rod-shaped bacterial strain, designated Gsoil 032T, was isolated from soil of a ginseng field in Pocheon Province, South Korea, and its taxonomic position was investigated by using a polyphasic approach. Strain Gsoil 032T grew at 10-42 °C and at pH 5.0-10.0 on R2A agar medium. Strain Gsoil 032T possessed ß-glucosidase activity, which was responsible for its ability to transform ginsenoside Rb1 (one of the dominant active components of ginseng) to compound K. On the basis of 16S rRNA gene sequence similarity, strain Gsoil 032T was shown to belong to the family Sphingobacteriaceae and to be related to Mucilaginibacter sabulilitoris SMS-12T (97.6 % sequence similarity) and Mucilaginibacter lappiensis ANJLI2T (97.1 %) The G+C content of the genomic DNA was 44.4 mol%. The predominant respiratory quinone was menaquinone MK-7 and the major fatty acids were summed feature 3 (comprising C16 : 1ω6c and/or C16 : 1ω7c), iso-C15 : 0 and iso-C17 : 0 3-OH. The major polar lipid detected was phosphatidylethanolamine, while the minor polar lipids were various unidentified aminophospholipids, unidentified phospholipids and unidentified polar lipids. DNA and chemotaxonomic data supported the affiliation of strain Gsoil 032T to the genus Mucilaginibacter. Strain Gsoil 032T could be differentiated genotypically and phenotypically from recognized species of the genus Mucilaginibacter. The isolate therefore represents a novel species, for which the name Mucilaginibacter pocheonensis sp. nov. is proposed, with the type strain Gsoil 032T (=KCTC 12641T=LMG 23495T).

Pseudoxanthomonas humi sp. nov., a bacterium isolated from rhizospheric soil of Fraxinus chinensis in Gyeonggi Province, South Korea.

A novel bacterial strain THG-MM13(T) was isolated from rhizospheric soil sample and was characterized by using a polyphasic approach. Cells were Gram-reaction-negative, non-motile and rod-shaped. The strain was aerobic, catalase and oxidase positive, and optimum growth temperature and pH were 28 °C and 7.0, respectively. On the basis of 16S rRNA gene sequence analysis, strain THG-MM13(T) (KM598260) belongs to the genus Pseudoxanthomonas and is most closely related to Pseudoxanthomonas wuyuanensis KCTC 23877(T) (97.4 %) (JN247803), followed by Pseudoxanthomonas koreensis KCTC 12208(T) (96.7 %) (AY550263) and Pseudoxanthomonas yeongjuensis KACC 11580(T) (96.7 %) (DQ438977). The DNA G + C content was 63.7 mol%, and the predominant respiratory quinone was ubiquinone-8. The major polar lipids were diphosphatidylglycerol and phosphatidylethanolamine. The major fatty acids were iso-C15:0 (31.3 %) and iso-C16:0 (19.3 %). The DNA-DNA relatedness value between strain THG-MM13(T) and P. wuyuanensis KCTC 23877(T) was below 50 %. The DNA-DNA hybridization result and results of the genotypic analysis in combination with chemotaxonomic and physiological data demonstrated that strain THG-MM13(T) represented a novel species within the genus Pseudoxanthomonas, for which the name Pseudoxanthomonas humi is proposed. The type strain is THG-MM13(T) (=KACC 18280(T) = CCTCC AB 2015122(T)).

Flavisolibacter ginsenosidimutans sp. nov., with ginsenoside-converting activity isolated from soil used for cultivating ginseng.

A Gram-reaction-negative, aerobic, non-motile and rod-shaped bacterial strain designated Gsoil 636T was isolated from soil of a ginseng cultivation field in Pocheon Province, South Korea and its taxonomic position was investigated using a polyphasic approach. Gsoil 636T grew at 18-30 °C and at pH 6.0-8.0 on R2A medium. Gsoil 636T possessed ß-glucosidase activity, which was responsible for its ability to transform ginsenoside Rb1 (ones of the dominant active components of ginseng) to F2. On the basis of 16S rRNA gene sequence similarity, Gsoil 636T was shown to belong to the family Chitinophagaceae and to be related to Flavisolibacter ginsengiterrae Gsoil 492T (96.7 % sequence similarity), Flavisolibacter ginsengisoli Gsoil 643T (96.6 %) and Flavisolibacter rigui 02SUJ3T (96.6 %). The G+C content of the genomic DNA was 48.9 %. The predominant respiratory quinone was MK-7 and the major fatty acids were iso-C15 : 0, summed feature 3 (comprising C16 : 1ω6c and/or C16 : 1ω7c) and iso-C17 : 0 3-OH. DNA and chemotaxonomic data supported the affiliation of Gsoil 636T to the genus Flavisolibacter. Gsoil 636T could be differentiated genotypically and phenotypically from the species of the genus Flavisolibacter with validly published names. The isolate therefore represents a novel species, for which the name Flavisolibacter ginsenosidimutans sp. nov. is proposed, with the type strain Gsoil 636T (KCTC 22818T = JCM 18197T = KACC 14277T).

Nocardioides ungokensis sp. nov., isolated from lake sediment.

A Gram-reaction-positive, aerobic, coccus- to rod-shaped, non-motile, non-spore-forming bacterium (strain UKS-03T) was isolated from a sediment sample of Ungok Lake in Gochang, Republic of Korea. The taxonomic position of this bacterium was determined in an investigation based on a polyphasic approach. On the basis of 16S rRNA gene sequence analysis, strain UKS-03T was shown to belong to the family Nocardioidaceae and to be related most closely to Nocardioides ginsengisegetis Gsoil 485T (98.5 % similarity), Nocardioides koreensis MSL-09T (98.4 %) and 'Nocardioides panaciterrulae' Gsoil 958 (97.3 %). Strain UKS-03T was characterized chemotaxonomically as having ll-2,6-diaminopimelic acid in its cell-wall peptidoglycan, MK-8(H4) as the predominant menaquinone, diphosphatidylglycerol and phosphatidylglycerol as the main polar lipids, and iso-C16 : 0, C17 : 1ω8c and C17 : 0 10-methyl as its major fatty acids. The G+C content of the genomic DNA was 71.9 mol%. Mean DNA-DNA relatedness values between strain UKS-03T and N. ginsengisegetis Gsoil 485T, N. koreensis KCTC 19272T and 'N. panaciterrulae' Gsoil 958 were 37.5 ± 7.2, 6.8 ± 0.9 and 3.1 ± 0.7 %, respectively. On the basis of the data from this polyphasic taxonomic study, strain UKS-03T represents a novel species of the genus Nocardioides, for which the name Nocardioides ungokensis sp. nov. is proposed. The type strain is UKS-03T ( = KACC 18304T = LMG 28591T).

Pseudoclavibacter terrae sp. nov. isolated from rhizosphere soil of Ophiopogon japonicus.

Strain THG-MD12T, a Gram-reaction-positive, aerobic, non-motile, rod-shaped bacterium was isolated from rhizosphere soil of Ophiopogon japonicus in PR China. THG-MD12T was closely related to members of the genus Pseudoclavibacter and showed the highest 16S rRNA gene sequence similarities with Pseudoclavibacter helvolus KCTC 19531T (98.8 %) and Pseudoclavibacter chungangensis KCTC 22691T (96.9 %). DNA-DNA hybridization showed 41.9 ± 2.1 % and 12.4 ± 0.9 % DNA reassociation with P. helvolus KCTC 19531T and P. chungangensis KCTC 22691T, respectively. Chemotaxonomic analyses revealed that strain THG-MD12T possesses menaquinone-9 as the predominant respiratory quinone, 2,4-diaminobutyric acid as the diamino acid in the peptidoglycan and anteiso-C15 : 0, iso-C16 : 0, C16 : 0 and anteiso-C17 : 0 as the major fatty acids. The polar lipid profile was found to consist of diphosphatidylglycerol, phosphatidylglycerol, two unknown glycolipids and two unknown lipids. These data corroborated the affiliation of THG-MD12T to the genus Pseudoclavibacter. Thus, the isolate represents a novel species, for which the name Pseudoclavibacter terrae sp. nov. is proposed, with THG-MD12T as the type strain ( = CCTCC AB 2015124T = KCTC 39562T).

Sphingobium soli sp. nov. isolated from rhizosphere soil of a rose.

Strain THG-SQA7(T), a Gram-negative, strictly aerobic, non-motile, rod-shaped bacterium was isolated from rhizosphere soil of a rose in PR China. Strain THG-SQA7(T) is closely related to the members of the genus Sphingobium, showing the highest 16S rRNA gene sequence similarities with Sphingobium lactosutens KACC 18100(T) (98.2%) and Sphingobium abikonense KCTC 2864(T) (98.1%). The DNA-DNA relatedness between strain THG-SQA7(T) and S. lactosutens KACC 18100(T) and S. abikonense KCTC 2864(T) was 26.2 ± 0.9 and 28.3 ± 1.2%, respectively. Chemotaxonomic data showed that strain THG-SQA7(T) possesses ubiquinone Q-10 as the predominant respiratory quinone, and C(18:1)ω7c, C(16:0), summed feature 3 (C(16:1)ω7c and/or C(16:1)ω6c) and C(14:0) 2OH as the major fatty acids. The major polar lipids were found to be phosphatidylethanolamine, phosphatidylmonomethylethanolamine, phosphatidylglycerol, sphingoglycolipid, diphosphatidylglycerol and phosphatidyldimethylethanolamine. Based on these results, together with phenotypic characterization, a novel species, Sphingobium soli sp. nov. is proposed.with the type strain is THG-SQA7(T) (=CCTCC AB 2015125(T) = KCTC 42607(T)).

20(S)-Ginsenoside Rh2 as aldose reductase inhibitor from Panax ginseng.

The root of Panax ginseng C. A. Meyer (Araliaceae) is a well-known herbal medicine in East Asia. The major bioactive metabolites in this root are commonly identified as ginsenosides. A series of ginsenosides were determined for in vitro human recombinant aldose reductase. This Letter aims to clarify the structural requirement for aldose reductase inhibition. We discovered that only ginsenoside 20(S)-Rh2 showed potent against aldose reductase, with an IC50 of 147.3 µM. These results implied that the stereochemistry of the hydroxyl group at C-20 may play an important role in aldose reductase inhibition. An understanding of these requirements is considered necessary in order to develop a new type of aldose reductase inhibitor. Furthermore, P. ginseng might be an important herbal medicine in preventing diabetic complications.

Preparation of przewalskinic acid A from salvianolic acid B using a crude enzyme from an Aspergillus oryzae strain.

Przewalskinic acid A is a rare, water-soluble, and highly biologically active ingredient found, thus far, only in the Salvia przewalskii Maxim herb; however, the content in S. przewalskii herb is very low. In order to obtain useful quantities of przewalskinic acid A, the biotransformatin of salvianolic acid B from Salvia miltiorrhiza root (danshen in Chinese) into przewalskinic acid A was studied using a crude enzyme produced from Aspergillus oryzae D30s strain. The crude enzyme from the A. oryzae strain hydrolyzed salvianolic acid B into przewalskinic acid A and danshensu. The preparation afforded 31.3 g przewalskinic acid A (91.0 % purity) and 13.1 g danshensu (95 % purity) from 75 g salvianolic acid B. The preparation of przewalskinic acid A was therefore very successful with a yield of over 86 %, but the yield of danshensu was only 33 %. The product przewalskinic acid A was identified using ultra-performance liquid chromatography-mass spectrometry (UPLC-MS) and NMR.

Obstacle classification and 3D measurement in unstructured environments based on ToF cameras.

Inspired by the human 3D visual perception system, we present an obstacle detection and classification method based on the use of Time-of-Flight (ToF) cameras for robotic navigation in unstructured environments. The ToF camera provides 3D sensing by capturing an image along with per-pixel 3D space information. Based on this valuable feature and human knowledge of navigation, the proposed method first removes irrelevant regions which do not affect robot's movement from the scene. In the second step, regions of interest are detected and clustered as possible obstacles using both 3D information and intensity image obtained by the ToF camera. Consequently, a multiple relevance vector machine (RVM) classifier is designed to classify obstacles into four possible classes based on the terrain traversability and geometrical features of the obstacles. Finally, experimental results in various unstructured environments are presented to verify the robustness and performance of the proposed approach. We have found that, compared with the existing obstacle recognition methods, the new approach is more accurate and efficient.

Full Point Encoding for Local Feature Aggregation in 3-D Point Clouds.

Point cloud processing methods exploit local point features and global context through aggregation which does not explicitly model the internal correlations between local and global features. To address this problem, we propose full point encoding which is applicable to convolution and transformer architectures. Specifically, we propose full point convolution (FuPConv) and full point transformer (FPTransformer) architectures. The key idea is to adaptively learn the weights from local and global geometric connections, where the connections are established through local and global correlation functions, respectively. FuPConv and FPTransformer simultaneously model the local and global geometric relationships as well as their internal correlations, demonstrating strong generalization ability and high performance. FuPConv is incorporated in classical hierarchical network architectures to achieve local and global shape-aware learning. In FPTransformer, we introduce full point position encoding in self-attention, that hierarchically encodes each point position in the global and local receptive field. We also propose a shape-aware downsampling block that takes into account the local shape and the global context. Experimental comparison to existing methods on benchmark datasets shows the efficacy of FuPConv and FPTransformer for semantic segmentation, object detection, classification, and normal estimation tasks. In particular, we achieve state-of-the-art semantic segmentation results of 76.8% mIoU on S3DIS sixfold and 73.1% on S3DIS Area 5. Our code is available at https://github.com/hnuhyuwa/FullPointTransformer.

Event-triggered adaptive neural prescribed performance admittance control for constrained robotic systems without velocity measurements.

In this paper, an event-triggered adaptive neural prescribed performance admittance control (ETANPPAC) scheme is proposed to control the constrained robotic systems without velocity sensors. To ensure compliance during human-robot interaction, the reference trajectory is obtained by reshaping the desired trajectory for the robotic systems based on the admittance relationship, where a saturation function is used to constrain the reference trajectory, avoiding excessive contact forces that could render the trajectory inexecutable. Moreover, a barrier Lyapunov function is used to constrain the tracking errors for prescribed performance, where a velocity observer and a radial basis function neural network are designed to estimate the velocity and the uncertainty of the robotic systems, respectively, to enhance control performance. To reduce the communication burden, an event-triggered mechanism is introduced and the Zeno behavior is avoided with the event-triggered condition. The stability of the whole control scheme is analyzed by the Lyapunov function. Simulation and experimental tests demonstrate that the proposed ETANPPAC scheme can track the desired trajectory well under constraints and reduce the communication burden, thereby achieving better efficiency for controlling the robotic systems compared with similar control schemes.

Protodioscin-glycosidase-1 hydrolyzing 26-O-ß-D-glucoside and 3-O-(1 → 4)-α-L-rhamnoside of steroidal saponins from Aspergillus oryzae.

A novel protodioscin-(steroidal saponin)-glycoside hydrolase, named protodioscin-glycosidase-1 (PGase-1), was purified and characterized from the Aspergillus oryzae strain. The molecular mass of this enzyme was determined to be about 55 kDa based on SDS-polyacrylamide gel electrophoresis. PGase-1 was able to hydrolyze the terminal 26-O-ß-D-glucopyranoside of protodioscin (furostanoside) to produce dioscin (spirostanoside), and then further hydrolyze the terminal 3-O-(1 → 4)-α-L-rhamnopyranoside of dioscin to form progenin III. However, PGase-1 could hardly hydrolyze the 3-O-(1 → 2)-α-L-rhamnopyranoside of progenin III, 3-O-ß-D-glucoside of trillin, and the 1-O-glycosides of ophiopogonin D (steroidal saponin). In addition, PGase-1 also could hydrolyze the α-D-galactopyranoside, ß-D-glucopyranoside, and ß-D-galactopyranoside of p-nitrophenyl-glycosides, but the enzyme could not hydrolyze the α-D-mannopyranoside, α-L-arabinopyranoside, α-D-glucopyranoside, ß-D-xylopyranoside, and α-L-rhamnopyranoside of p-nitrophenyl-glycosides. These new properties of PGase-1 were significantly different from those of previously described steroidal saponin-glycosidases and the glycosidases currently described in Enzyme Nomenclature by the NC-IUBMB. The gene (termed as pgase-1) encoding PGase-1 was cloned, sequenced, and expressed in Pichia pastoris GS115. The complete nucleotide sequence of pgase-1 consists of 1,725 bp. The recombinant PGase-1 from recombinant P. pastoris GS115 strain also showed the activity hydrolyzing glycosides of steroidal saponins which was similar to that of the wild-type PGase-1 from A. oryzae. The PGase-1 gene is highly similar to Aspergilli α-amylase (EC 3.2.1.1), and PGase-1 should be classified as glycoside hydrolase family 13 by the method of gene sequence-based classification. But the enzyme properties of PGase-1 are different from those of α-amylase in this family.

Preparation of progenin III from total steroidal saponins of Dioscorea nipponica Makino using a crude enzyme from Aspergillus oryzae strain.

Progenin III, one of the most active spirostanol saponins, is a potential candidate for anti-cancer therapy due to its strong antitumor activity and low hemolytic activity. However, the concentration of progenin III is extremely low in natural Dioscorea plants. In this paper, the progenin III production from total steroidal saponins of Dioscorea nipponica Makino was studied using the crude enzyme from Aspergillus oryzae DLFCC-38. The crude enzyme converting total steroidal saponins into progenin III was obtained from the A. oryzae DLFCC-38 culture. For enzyme production, the strain was cultured for 72 h at 30 °C with shaking at 150 rpm in 5 % (w/v) malt extract medium containing 2 % (v/v) extract of D. nipponica as the enzyme inducer. The crude enzyme converted total steroidal saponins into major progenin III with a high yield when the reaction was carried out for 9 h at 50 °C and pH 5.0 with the 20 mg/ml of substrate. In the preparation of progenin III, 117 g of crude progenin III was obtained from 160 g of substrate, and the crude product was purified with silica gel column to obtain 60.3 g progenin III of 93.4 % purity.

Aspergillus niger DLFCC-90 rhamnoside hydrolase, a new type of flavonoid glycoside hydrolase.

A novel rutin-α-L-rhamnosidase hydrolyzing α-L-rhamnoside of rutin, naringin, and hesperidin was purified and characterized from Aspergillus niger DLFCC-90, and the gene encoding this enzyme, which is highly homologous to the α-amylase gene, was cloned and expressed in Pichia pastoris GS115. The novel enzyme was classified in glycoside-hydrolase (GH) family 13.

Fully Convolutional Network-Based Self-Supervised Learning for Semantic Segmentation.

Although deep learning has achieved great success in many computer vision tasks, its performance relies on the availability of large datasets with densely annotated samples. Such datasets are difficult and expensive to obtain. In this article, we focus on the problem of learning representation from unlabeled data for semantic segmentation. Inspired by two patch-based methods, we develop a novel self-supervised learning framework by formulating the jigsaw puzzle problem as a patch-wise classification problem and solving it with a fully convolutional network. By learning to solve a jigsaw puzzle comprising 25 patches and transferring the learned features to semantic segmentation task, we achieve a 5.8% point improvement on the Cityscapes dataset over the baseline model initialized from random values. It is noted that we use only about 1/6 training images of Cityscapes in our experiment, which is designed to imitate the real cases where fully annotated images are usually limited to a small number. We also show that our self-supervised learning method can be applied to different datasets and models. In particular, we achieved competitive performance with the state-of-the-art methods on the PASCAL VOC2012 dataset using significantly fewer time costs on pretraining.