LeaderBoard-CVWC2019

Test-dev
Team	mAP	BFLOPs	PPF-b	PPF-a
[1]AlexNet in my heart	0.515	0.0712	1.199	0.675
[2]dcyhw	0.586	5.68	0.027	0.519
[3]Team Galactic	0.479	0.466	0.106	0.405
[4]zdi	0.601	245.3	0.001	0.388
[5]lazy-learners	0.546	112.48	0.001	0.388
[6]YuTeam	0.492	7.5	0.008	0.343
CVWC SSDv2	0.476	1.25	0.037	0.304
[7]DeepBlueAI	0.442	0.25	0.046	0.194
CVWC SSDv1	0.426	1.2	-0.003	0.125
[8]bagh	0.262	0.00546	-30.776	0.111
[9]eyeoftiger	0.456	535	0.000	0.104
[10]DelPro	0.480	N/A	N/A	0
[11]Bestfitting_NBU	0.460	N/A	N/A	0
[12]tony19	0.507	N/A	N/A	0
[13],	0.465	N/A	N/A	0
[14]TeamEVAC	0.014	N/A	N/A	0

Full Test
Team	mAP	BFLOPs	PPF-b	PPF-a
[1]AlexNet in my heart	0.513	0.0712	1.169	0.675
[2]dcyhw	0.586	5.680	0.028	0.519
[3]Team Galactic	0.475	0.466	0.097	0.405
[4]zdi	0.601	245.3	0.001	0.388
[5]lazy-learners	0.548	112.48	0.001	0.388
[6]YuTeam	0.482	7.5	0.007	0.311
CVWC SSDv2	0.473	1.25	0.034	0.304
[7]DeepBlueAI	0.432	0.25	0.009	0.194
CVWC SSDv1	0.418	1.2	-0.010	0.125
[8]bagh	0.259	0.00546	-31.248	0.111
[9]eyeoftiger	0.459	535	0.000	0.104
[10]DelPro	0.480	N/A	N/A	0
[11]Bestfitting_NBU	0.460	N/A	N/A	0
[12]tony19	0.507	N/A	N/A	0
[13],	0.465	N/A	N/A	0
[14]TeamEVAC	0.014	N/A	N/A	0

Team	Members	Repo	Description
[1]AlexNet in my heart	heart/Batiary Orest Kupyn(Ukrainian Catholic University), Dzmitry Pranchuk(WANNABY), Volodymyr Budzan	https://github.com/KupynOrest/AmurTigerCVWC	We used custom network architecture based on RetinaNet with MobileNetv1 backbone. To achieve greated efficiency we use Depthwise Separable Convolutions instead of regular one and smaller num of layers in localization/classification head. We train our net on 224 img size. As the second stage we also use knowledge distillation via pseudolabelling with large SE-ResNeXt-101 network.
[2]dcyhw	Jianwei Dong, Longrong Yang, Junjie Chen, Junqiang Huang, Xinxu Wei	https://github.com/JwDong2019/CVWC-2019-FCOS	We solve the detection track by adopting FCOS which is proposed in CVPR 2019 and our backbone is resnet 50. We used COCO datasets for pretraining and deformable convolution is also added into our method for enhanced performance.
[3]Team Galactic	Wei Runchen, Luo Dawei, Yin Xiaojie	https://github.com/RunchenWei/YOLO-mini-tiger	1.We use YOLOv3 method based on Darknet framework, and its backbone is EfficientNet_b0( its structure have been adjusted to single label classification with small sample dataset.) The loss function is replaced with Generalized Intersection over Union Loss(GIoU Loss). 2.All of our submission was trained on Amur Tiger trainset without any tricks (without multi-scales or flip in traning and evaluation) and We use Gaussian distribution random initialization without any pretrained model. 3. We propose a new backbone structure adjustment algorithm to reduce the BFLOPs and model size in any backbone.( It is only used in single label classification with small sample dataset, it is different from prune algorithm ).
[4]zdi	Zonglin Di(TJU), Bingchen Zhao(TJU)	https://github.com/ElegantLin/CVWC-2019	The base net I used is Faster RCNN using HRNet as backbone from https://github.com/HRNet/HRNet-Object-Detection. The pretrained model is from ImageNet and I fine-tuned it with training set. We tried several tricks and the best mAP we got is 0.60009 with 245.3 GFLOPs.
[5]lazy-learners	Srinivas VARADHARAJAN (Dell EMC), Vish KAVASSERYRAJALINGAM (Samsung)	SRINIVAS.VARADHARAJAN@live.com viswa1410@gmail.com	The model is based on ResNet feature extractor pre-trained on MS-COCO dataset, the detection head is a FasterRCNN based model. There were several data augmentations technique added to augment the training data size. Hyperparameter tuning was effectively done after multiple experiments.
[6]YuTeam	Yuanqiang Cai, Libo Zhang(ISCAS), Dawei Du	https://github.com/yuchengtianxia/SLN_Final	We present a tiger detector, SLN(Final). The SLN is an anchor-free tiger detector. Its backbone network is MobileNetV1. It uses FPN and Deformable Convolution to fuse and modify features. It has light FLOPs (7.5B), training parameters(4,505,816), and model size (.data(72.2MB)+ .index(21.7KB)+.meta(3.3MB)).
[7]DeepBlueAI	Zhipeng Luo, Feng Ni, Zhenyu Xu, Yuehan yao, Bin Dong	https://github.com/nemonameless/mmdettiger	We adopts MobileNetV2-SSDLite, achieving the trade-off between mAP and FLOPs by reducing 50% number of channels. We only use ImageNet pre-trained MobileNetV2 model. And we do not use multiple models, multi-scales or flip in the evaluation, just single model and single scale(300*300) for training and testing. We just use soft NMS for a little gain.
[8]bagh	saket kunwar	https://github.com/saketkunwar/cvwc2019_det	Model is yolov3-tiny with custom anchors determined from ground truth boxes. Pytorch was used for training using coco pretrained weights. But for inference the trained model was converted to intel openvino format with fixes for proper image re-scaling and anchor sizes. Yolov3 was also tested with pytorch and openvino but final submitted result on leader-board is yolov3-tiny.
[9]eyeoftiger	Anay Majee(Intel), Kshitij Agrawal(Intel)	https://github.com/kshitijagrwl/tiger_detection	SSD-inceptionv2 We have used a vanila SSD network with an inception v2 backbone at a resolution of 300x300. The model is pretrained on COCO Detection task and finetuned for the ATRW dataset.
[10]DelPro	-
[11]Bestfitting_NBU	-
[12]tony19	nemonameless@qq.com
[13],	hamletkong@gmail.com		No Response
[14]TeamEVAC			No Response

Test-dev
Team	mAP
[1]DeepBlueAI	0.904
[2]UCAS-ISCAS	0.900
[3]eyeoftiger	0.891
CVWC HRNet	0.876
[4]DelPro	0.867
[5]bagh	0.867
[6]CV_tiger	0.838
[7]AlexNet in my heart	0.741
[8]naive_baseline	0.678
CVWC AlphaPose	0.576

Full Test
Team	mAP
[1]DeepBlueAI	0.894
[2]UCAS-ISCAS	0.892
[3]eyeoftiger	0.874
CVWC HRNet	0.869
[4]DelPro	0.859
[5]bagh	0.849
[6]CV_tiger	0.835
[7]AlexNet in my heart	0.732
[8]naive_baseline	0.671
CVWC AlphaPose	0.574

Team	Members	Repo	Description
[1]DeepBlueAI	Zhipeng Luo, Feng Ni, Zhenyu Xu, Yuehan yao, Bin Dong	https://github.com/dbofseuofhust/Pose-Track-Solutions	We chose posenet as the basic algorithm, and improved UNET,HRNET as our backbone and added spatial attentions to the downsampling layers to obtain more spatial information. Additional data augments are made to the data set to improve generalization ability. We use COCO pre-trained models.
[2]UCAS-ISCAS	Hao Wang(UCAS), Libo Zhang(ISCAS), Longyin Wen	https://github.com/JisuiWang/CVWC2019-pose	Our Solution is mostly based on the CVPR19 work: Deep High-Resolution Representation Learning for Human Pose Estimation, and we just modify some parameters in the open-sourced code according to the difference of keypoints between human and tiger. We used the pretrained model as the paper described which is trained on imagenet.
[3]eyeoftiger	Anay Majee(Intel), Kshitij Agrawal(Intel)	https://github.com/amajee11us/ATRW_Ensemble_HRnet	The key aspect of our approach was to improve on the corner cases on already effective HRNet network. For this reason, the following methodology was adopted: 1.We conducted experiments with multi resolution images to test the effect of resolution on the model. We finally settled on 640x480 input size. 2.During training we adopted a 5 fold split on the entire train+validation dataset. 3.For improving the accuracy during inference, the 5 fold split was ensembled using multiple approaches - average ensemble, bagging ensemble and random forest ensemble for obtaining the best results from the solution. For the submission we have selected average ensemble as it performed the best in our experiments. 4.All the models used we trained on HRNet-W32 network which were pre-trained on ImageNet dataset.
[4]DelPro	Ankita Shukla(IIITD), Gullal Singh Cheema(IIITD), Connor Anderson(BYU), Pei Gao(BYU), Ryan Farrell(BYU), Saket Anand(IIITD), Divyam Anshumaan(IIITD)	https://github.com/FGVC/DelPro	We use an HRNet-W48 model, pretrained on the COCO dataset. We use data augmentation like random rotation and random horizontal flipping. We train on images resized to 384x288.
[5]bagh	saket kunwar		Hrnet was used in this track with image dimension of 288x384. Imagenet pretrained weights were used.
[6]CV_tiger	1564480490@qq.com
[7]AlexNet in my heart	heart/Batiary Orest Kupyn(Ukrainian Catholic University), Dzmitry Pranchuk(WANNABY), Volodymyr Budzan	https://github.com/KupynOrest/AmurTigerCVWC	Simple FPN network with SE-ResNeXt-50 backbone for heatmap prediction. Each keypoint is predicted as the saparate heatmap using soft IoU loss.
[8]naive_baseline	me@arseny.info		No Response

Test-dev
Team	mmAP	Single Cam			Cross Cam
Team	mmAP	mAP	top-1	top-5	mAP	top-1	top-5
[1]Bestfitting_NBU	0.812	0.909	0.986	0.997	0.715	0.937	0.971
[2]BRL-RedPanda	0.773	0.911	0.986	0.989	0.635	0.903	0.971
[3]NWPU_ASGO	0.744	0.879	0.977	0.989	0.609	0.914	0.937
[4]DeepBlueAI	0.695	0.866	0.977	0.986	0.525	0.857	0.909
[5]DelPro	0.691	0.847	0.986	0.986	0.535	0.891	0.949
[6]SDL	0.672	0.860	0.949	0.969	0.484	0.783	0.874
[7]AI-TIGER	0.665	0.848	0.949	0.963	0.482	0.783	0.869
[9]NDWild	0.663	0.777	0.923	0.974	0.550	0.846	0.977
[8]zdi	0.649	0.842	0.954	0.986	0.457	0.869	0.926
[11]Batiary	0.636	0.766	0.923	0.980	0.506	0.823	0.960
[10]pr	0.633	0.816	0.940	0.966	0.451	0.783	0.874
[12]aaa	0.623	0.763	0.920	0.977	0.482	0.771	0.891
CVWC Triplet	0.598	0.714	0.866	0.954	0.481	0.794	0.937
[14]DiKorsch	0.546	0.706	0.869	0.951	0.386	0.731	0.869
[13]bagh	0.545	0.636	0.837	0.926	0.455	0.777	0.909
[15]ThisIsMyTeam	0.532	0.674	0.860	0.969	0.390	0.834	0.914
[16]cavs	-	-	-	-	-	-	-

Full Test
Team	mmAP	Single Cam			Cross Cam
Team	mmAP	mAP	top-1	top-5	mAP	top-1	top-5
[1]Bestfitting_NBU	0.816	0.906	0.977	0.991	0.726	0.936	0.967
[2]BRL-RedPanda	0.770	0.898	0.966	0.977	0.643	0.913	0.958
[3]NWPU_ASGO	0.751	0.879	0.969	0.983	0.622	0.925	0.951
[4]DeepBlueAI	0.704	0.865	0.956	0.983	0.543	0.889	0.929
[5]DelPro	0.696	0.836	0.973	0.981	0.556	0.872	0.948
[6]SDL	0.672	0.857	0.940	0.960	0.488	0.783	0.867
[7]AI-TIGER	-	-	-	-	-	-	-
[8]zdi	0.658	0.846	0.954	0.984	0.470	0.841	0.904
[9]NDWild	0.658	0.763	0.907	0.967	0.553	0.851	0.944
[10]pr	-	-	-	-	-	-	-
[11]Batiary	0.634	0.757	0.900	0.967	0.511	0.824	0.935
[12]aaa	0.631	0.758	0.906	0.964	0.505	0.806	0.917
CVWC Triplet	0.592	0.713	0.866	0.960	0.471	0.775	0.905
[13]bagh	0.549	0.650	0.852	0.942	0.448	0.769	0.900
[14]DiKorsch	0.542	0.694	0.854	0.946	0.391	0.742	0.871
[15]ThisIsMyTeam	0.537	0.682	0.869	0.967	0.392	0.811	0.914
[16]cavs	-	-	-	-	-	-	-

Team	Members	Repo	Description
[1]Bestfitting_NBU	Cen Liu, Lijun Guo(NBU), Rong Zhang(NBU), Wenchao Chang	https://github.com/LcenArthas/CWCV2019-Amur-Tiger-Re-ID	We introduce a novel triple-stream framework for the tiger Re-ID challenge, which consists of a full-image stream, a body-image stream and a limbs- image stream. Specifically, in the train stage, considering the missing and the errors of keypoint annotations, the body and limbs streams are used as regulators to guide the original full image stream in learning and aligning the local features. In the inference stage, we only use the learnt full image stream, which performs well with high efficiency and accuracy in the tiger Re-ID task, especially in the wild task (since it doesn’t need the pose estimation step).
[2]BRL-RedPanda	Ning Liu, Nan Zhang, Xinhua Cheng, Jianing Zhu, Qijun Zhao	https://github.com/liuning-scu-cn/AmurTigerReID	Our solution uses SE-ResNet50 model as backbone which was pre-trained by ILSVRC. In addition, we design two complementary network branches to learn multiple discriminative features. We use multi-task learning strategy to supervise the model training. Finally, we fine-tune the model with triplet loss. The Re-ID results are obtained based on the fusion of the learned multiple features.
[3]NWPU_ASGO	Jiwen Yu, Junnan Liu, Zhizheng Yang, Zhouyangzi Zhang, Yixin Zhu, Haibo Su, Lu Yang, Bingliang Jiao, Peng Wang	https://github.com/vvictoryuki/tiger_reid_pytorch	We extract global features and local features of images on CNN, optimize these features with Triplet loss and id classification loss and apply several methods for data augmentation such as flip as new id, random whitening, random crop and so on. Besides, we proposed an example sampling strategy for training using hard negative mining. Finally, we ensemble our models with different backbones and epochs using imagenet pre-trained models (ResNet50, ResNet152, DenseNet161, DenseNet169, DenseNet201, DenseNet121) downloaded from pytorch.
[4]DeepBlueAI	Zhipeng Luo, Feng Ni, Zhenyu Xu, Yuehan yao, Bin Dong	https://github.com/John-Yao/CVWC2019_ReID	We chose posenet as the basic algorithm, and improved UNET,HRNET as our backbone and added spatial attentions to the downsampling layers to obtain more spatial information. Additional data augments are made to the data set to improve generalization ability. We use COCO pre-trained models.
[5]DelPro	Ankita Shukla(IIITD), Gullal Singh Cheema(IIITD), Connor Anderson(BYU), Pei Gao(BYU), Ryan Farrell(BYU), Saket Anand(IIITD), Divyam Anshumaan(IIITD)	https://github.com/FGVC/DelPro	We used pretrained DenseNet 121 on ImageNet dataset. The model is finetuned with a loss function defined on class probabilities of similar and dissimilar pairs along with the cross entropy loss and with a set of input transformations. During testing, initial ranking is obtained based on similarities in the feature representation. Followed by using the left and right flank information obtained using key points and rearranging the ranking matrix with SIFT matching.
[6]SDL	Nan Jiang(UCAS), Yuqi Gong(UCAS), Xuehui Yu(UCAS), Zhenjun Han(UCAS)	https://github.com/Vvfshadow/feature-aggregation	Our method is simple and straight-forward: we use ResNet-50 (pretrained on ImageNet) as backbone, and funetune it by using the given training dataset only. To effectively extract part features, we utilize two strategies: 1) uniformly partition in the feature maps; 2) use pose information to divide feature maps. During testing, we use two models for ensemble and adopt query expansion and rerank for post-processing.
[7]AI-TIGER	815719905@qq.com		No Response
[8]zdi	Zonglin Di(TJU), Bingchen Zhao(TJU)	https://github.com/ElegantLin/CVWC-2019	We train our reID model using ImageNet pretrained ResNet50 as backbone, add a BNNeck at the feature, combined with Warmup lr scheduler and random erasing data augmentation. Only global feature is used for our reID model, and no pose information is added.
[9]NDWild	Bryan (Ning) Xia(University of Notre Dame), Prof. Christian Poellabauer(University of Notre Dame)	https://github.com/baxincs/ewan-reid	Our solution is based on the paper "Non-local Second-order Attention Networks for Person Re-identification”. The network architecture is modified based on ResNet50, therefore, we initialize our network with the pre-trained weights on ImageNet.
[10]pr	1007284540@qq.com		No Response
[11]AlexNet in my heart/Batiary	Orest Kupyn(Ukrainian Catholic University), Dzmitry Pranchuk(WANNABY), Volodymyr Budzan	https://github.com/KupynOrest/AmurTigerCVWC	For this task we use custom architecture with SE-ResNext backbone. The last layers and loss function are taken from CosFace - https://arxiv.org/abs/1801.09414 paper. (BN->Droput->BN1d->Dense) . The Dense layer output the final embedding of size = 2048. We use Large Margin Cosine Loss to train the net to distinguish between different tigers. For Re-ID we compute Cosine Distance between embeddings and rank them accordingly. The final submit is an ensemble of 4 models train of different splits/augmentations. The ensemble is based on majority voting.
[12]aaa	gullalcheema@gmail.com ashukl20@asu.edu		No Response
[13]bagh	-
[14]DiKorsch	dimitri.korsch@uni-jena.de		No Response
[15]ThisIsMyTeam	korschdima@gmail.com		No Response
[16]cavs	christopher_my@163.com		No Response

Test-dev
Team	mmAP	Single Cam			Cross Cam
Team	mmAP	mAP	top-1	top-5	mAP	top-1	top-5
[1]Bestfitting_NBU	0.805	0.880	0.962	0.973	0.729	0.914	0.953
[2]DeepBlueAI	0.690	0.841	0.944	0.964	0.539	0.893	0.930
[3]Batiary	0.663	0.785	0.896	0.960	0.540	0.844	0.930
[4]zdi	0.640	0.820	0.931	0.958	0.459	0.838	0.904
CVWC Baseline	0.586	0.711	0.835	0.902	0.461	0.766	0.886

Full Test
Team	mmAP	Single Cam			Cross Cam
Team	mmAP	mAP	top-1	top-5	mAP	top-1	top-5
[1]Bestfitting_NBU	0.807	0.889	0.956	0.974	0.724	0.929	0.956
[2]DeepBlueAI	0.696	0.845	0.944	0.967	0.548	0.908	0.941
[3]Batiary	0.666	0.789	0.909	0.959	0.543	0.856	0.942
[4]zdi	0.644	0.823	0.932	0.967	0.465	0.849	0.910
CVWC Baseline	0.582	0.703	0.827	0.907	0.460	0.778	0.882

Team	Members	Repo	Description
[1]Bestfitting_NBU	Cen Liu, Lijun Guo(NBU), Rong Zhang(NBU), Wenchao Chang	https://github.com/LcenArthas/CWCV2019-Amur-Tiger-Re-ID	We use the Mask RCNN_50ResNet as the detector and use the proposed network in the plain Re-ID task as tracker for the wild re-id task. Especially, our method does not need keypoint information in the inference stage and is more suitable for the wild Re_id task.
[2]DeepBlueAI	Zhipeng Luo, Feng Ni, Zhenyu Xu, Yuehan yao, Bin Dong	https://github.com/John-Yao/CVWC2019_ReID	Without preprocessing of data, we build a ReID model base on global features and batch hard mining triplet loss. More details, some training tricks are used, including warmup,random erasing,centerloss and etc. For post-process, we use multi-scales augmentation and reranking. We use COCO pre-trained models.
[3]AlexNet in my heart/Batiary	Orest Kupyn(Ukrainian Catholic University), Dzmitry Pranchuk(WANNABY), Volodymyr Budzan	https://github.com/KupynOrest/AmurTigerCVWC	For this task we use custom architecture with SE-ResNext backbone. The last layers and loss function are taken from CosFace - https://arxiv.org/abs/1801.09414 paper. (BN->Droput->BN1d->Dense) . The Dense layer output the final embedding of size = 2048. We use Large Margin Cosine Loss to train the net to distinguish between different tigers. For Re-ID we compute Cosine Distance between embeddings and rank them accordingly. The final submit is an ensemble of 4 models train of different splits/augmentations. The ensemble is based on majority voting.
[4]zdi	Zonglin Di(TJU), Bingchen Zhao(TJU)	https://github.com/ElegantLin/CVWC-2019	For wild ReID track, We choose our best detection submit as reID gallery, crop the bounding boxes from the images and perform reID. The following steps are similar to those in Plain reID Track.