Overview

We offer a benchmark suite together with an evaluation server, such that authors can upload their results and get a ranking. We offer a Dataset that contains more than 50000 pictures, including 30462 images for training set, 10000 images for validation set and 10000 images for test set. If you would like to submit your results, please register, login, and follow the instructions on our submission page.

Note: We only display the highest submission of each person.

Single-Person Human Parsing Track

Metrics

We use four metrics from common semantic segmentation and scene parsing evaluations that are variations on pixel accuracy and region intersection over union (IU). The metrics are reported by FCN. The four metrics are Pixel accuracy(%) , Mean accuracy(%), Mean IoU(%) and Frequency weighted IoU(%). The details show per-class IoU(%).

User Id	Method	Pixel accuracy	Mean accuracy	Mean IoU	Frequency weighted IoU	Details	Abbreviation	Submit Time
258	AttEdgeNet	87.40	67.17	54.17	78.55	Details background hat hair glove sunglasses upper-clothes dress coat socks pants 87.66 66.22 72.37 42.81 28.04 70.66 40.00 58.60 44.72 75.14 jumpsuits scarf skirt face left-arm right-arm left-leg right-leg left-shoe right-shoe 29.67 25.05 30.45 74.12 64.12 65.69 57.94 57.72 46.64 45.83	Abbreviation Contributors Description Li Tianpeng, Chen Jiansheng Tsinghua University We added two new classes(Background on human body and 3pixel edges) when we trained our models. Two attention blocks were added to the resnet of pspnet to combine the low level and high level features. Finally we used the predicted coarse result as attention map multiply the Conv3 feature and then to train a 8-layers subnet to generate more accurate results. Our single model can achieve MeanIOU 52.6. Final results is generated by an ensemble of five models.	2018-04-09 04:28:56
284	Attention	84.52	54.83	44.60	74.03	Details background hat hair glove sunglasses upper-clothes dress coat socks pants 84.90 59.11 68.25 31.48 20.76 65.47 30.71 53.47 34.67 68.47 jumpsuits scarf skirt face left-arm right-arm left-leg right-leg left-shoe right-shoe 22.75 14.97 22.35 71.89 52.54 55.31 39.77 39.67 27.60 27.85	Abbreviation Contributors Description WuTao ABC	2018-05-03 15:45:48
287	n_v3	86.73	61.93	51.53	77.30	Details background hat hair glove sunglasses upper-clothes dress coat socks pants 86.75 63.59 70.81 36.70 22.68 68.55 37.10 57.35 38.33 73.04 jumpsuits scarf skirt face left-arm right-arm left-leg right-leg left-shoe right-shoe 28.09 12.35 28.37 74.33 62.98 64.23 58.71 58.89 44.32 43.40	Abbreviation Contributors Description	2018-06-04 01:58:10
283	JD_BUPT	87.42	65.86	54.44	78.34	Details background hat hair glove sunglasses upper-clothes dress coat socks pants 87.26 65.10 72.21 42.71 31.03 70.53 42.04 58.95 42.59 74.47 jumpsuits scarf skirt face left-arm right-arm left-leg right-leg left-shoe right-shoe 32.27 22.03 33.40 74.93 63.82 65.66 59.68 59.92 45.37 44.84	Abbreviation Contributors Description Xinchen Liu (BUPT&JD), Meng Zhang (BUPT), Yanan Li (BUPT), Wenhui Gao (BUPT), Jiangtian Pan (JD AI Research), Wu Liu (JD AI Research), Huadong Ma (BUPT) (1) We revised and finetuned the JPP-Net[1], SS-NAN[2], SSL[3], DenseNet[4], RefineNet[5] on LIP training set. Then we combined the five models with different fusion strategies. (2) We mined several hard classes to improve the overall performance. (3) We used the data augmentation, focal loss[6], and image morphology. [1]Liang et al., Look into Person: Joint Body Parsing & Pose Estimation Network and A New Benchmark, T-PAMI, 2018. [2]Zhao et al., Self-Supervised Neural Aggregation Networks for Human Parsing, CVPR Workshop, 2017. [3]Gong et al., Look into Person: Self-supervised Structure-sensitive Learning and A New Benchmark for Human Parsing, CVPR, 2017. [4]Jegou et al., The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation, CVPR, 2017. [5]Lin et al., RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation, CVPR, 2017. [6]Focal Loss for Dense Object Detection, Lin et al., ICCV, 2017.	2018-06-10 12:59:11
290	densenet&deeplabv3+	81.56	49.56	37.92	70.65	Details background hat hair glove sunglasses upper-clothes dress coat socks pants 83.19 50.91 63.94 15.78 4.58 59.92 26.81 47.19 27.71 62.74 jumpsuits scarf skirt face left-arm right-arm left-leg right-leg left-shoe right-shoe 17.66 2.88 11.60 68.13 44.82 48.30 36.70 36.21 25.92 23.40	Abbreviation Contributors Description hanqiuyuan I replace the backbone of Xception with densenet in the deeplabv3+ model.Meanwhile, I remove the last layer of densenet and the maxpool layer to get the feature map. The atrous I use are 3,6,9 and the output_stride is 32.The crop_size is 640 because the maxsize of the dataset id 640. Finally, I concat the Unit1 of Block2 with the resized and concated feature maps.	2018-05-27 05:33:01
292	no-ssl clean data 10w	83.73	53.13	42.69	73.04	Details background hat hair glove sunglasses upper-clothes dress coat socks pants 84.52 57.40 67.45 29.90 19.55 63.66 27.40 51.57 32.40 66.94 jumpsuits scarf skirt face left-arm right-arm left-leg right-leg left-shoe right-shoe 21.71 13.09 20.17 70.73 49.76 52.45 36.85 36.72 26.20 25.30	Abbreviation Contributors Description	2018-05-30 07:46:04
297	xNet	80.98	47.50	36.20	70.20	Details background hat hair glove sunglasses upper-clothes dress coat socks pants 84.03 61.58 69.31 29.35 22.69 58.93 27.68 45.39 31.34 65.49 jumpsuits scarf skirt face left-arm right-arm left-leg right-leg left-shoe right-shoe 8.25 9.27 19.56 72.61 2.73 38.18 12.44 32.50 13.34 19.33	Abbreviation Contributors Description	2018-06-06 04:21:49
301	refine_net	87.42	65.86	54.44	78.34	Details background hat hair glove sunglasses upper-clothes dress coat socks pants 87.26 65.10 72.21 42.71 31.03 70.53 42.04 58.95 42.59 74.47 jumpsuits scarf skirt face left-arm right-arm left-leg right-leg left-shoe right-shoe 32.27 22.03 33.40 74.93 63.82 65.66 59.68 59.92 45.37 44.84	Abbreviation Contributors Description	2018-06-10 13:02:27
270	PSPse	88.92	67.78	57.90	80.59	Details background hat hair glove sunglasses upper-clothes dress coat socks pants 89.05 68.88 75.38 46.98 36.80 73.11 38.31 60.97 50.19 77.11 jumpsuits scarf skirt face left-arm right-arm left-leg right-leg left-shoe right-shoe 29.25 26.10 30.89 77.87 69.74 71.15 66.43 66.33 51.80 51.76	Abbreviation Contributors Description Ting Liu(BJTU)*, Tao Ruan(BJTU)*, Yunchao Wei(UIUC), Shikui Wei(BJTU), Jinjun Xiong(UIUC, IBM), Yao Zhao(BJTU), Thomas Huang(UIUC). (* means equal contribution) We proposed a novel PSPse (‘se’ means segmentation with edge) network, which consists of three key modules to learn for parsing in an end-to-end manner: 1) high resolution embedding module; 2) global context embedding module; 3) edge perceiving module. The model is trained by sync batch norm across multiple GPUs. With the Res101 as the backbone, one single PSPse model can achieve the mIoU score of 56.5% without any bells and whistles.  Our best result is produced by an ensemble of three models. Ref: [1] Pyramid Scene Parsing Network. CVPR 2017  [2] Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. Arxiv 2018 [3] Magic-wall: Visualizing Room Decoration. ACM MM 2017.	2018-06-10 15:52:56
285		81.12	47.20	36.97	69.68	Details background hat hair glove sunglasses upper-clothes dress coat socks pants 82.62 56.63 64.52 30.96 26.47 61.50 24.97 49.33 35.64 64.00 jumpsuits scarf skirt face left-arm right-arm left-leg right-leg left-shoe right-shoe 18.96 6.60 19.37 68.26 16.30 29.89 21.25 22.47 21.92 17.80	Abbreviation Contributors Description	2018-06-07 07:14:03

---