GitPlanet

Cheap and reliable Node.js hosting starts at $3/month, and $1/month static HTML hosting

Created with love in Canada, visit hostnodejs.com today

Feel like to post an Ad? Learn Details
All Projects → yizt → keras-faster-rcnn

yizt / keras-faster-rcnn

Licence: other
keras实现faster rcnn，end2end训练、预测; 持续更新中,见todo... ;欢迎试用、关注并反馈问题

Programming Languages

python
139335 projects - #7 most used programming language

Labels

kerasfaster-rcnnobject-detectionr-cnn

Projects that are alternatives of or similar to keras-faster-rcnn

Caffe Faster Rcnn
faster rcnn c++ version. joint train; please checkout into dev branch (git checkout dev)
Stars: ✭ 210 (+147.06%)
Mutual labels:  faster-rcnn
gluon-faster-rcnn
Faster R-CNN implementation with MXNet Gluon API
Stars: ✭ 31 (-63.53%)
Mutual labels:  faster-rcnn
smd
Simple mmdetection CPU inference
Stars: ✭ 27 (-68.24%)
Mutual labels:  faster-rcnn
Syndata Generation
Code used to generate synthetic scenes and bounding box annotations for object detection. This was used to generate data used in the Cut, Paste and Learn paper
Stars: ✭ 214 (+151.76%)
Mutual labels:  faster-rcnn
Rrpn faster Rcnn tensorflow
A tensorflow re-implementation of RRPN: Arbitrary-Oriented Scene Text Detection via Rotation Proposals.
Stars: ✭ 243 (+185.88%)
Mutual labels:  faster-rcnn
publications-arruda-ijcnn-2019
Cross-Domain Car Detection Using Unsupervised Image-to-Image Translation: From Day to Night
Stars: ✭ 59 (-30.59%)
Mutual labels:  faster-rcnn
Traffic Sign Detection
Traffic Sign Detection. Code for the paper entitled "Evaluation of deep neural networks for traffic sign detection systems".
Stars: ✭ 200 (+135.29%)
Mutual labels:  faster-rcnn
Object-and-Semantic-Part-Detection-pyTorch
Joint detection of Object and its Semantic parts using Attention-based Feature Fusion on PASCAL Parts 2010 dataset
Stars: ✭ 18 (-78.82%)
Mutual labels:  faster-rcnn
Pytorch Faster Rcnn
pytorch based implementation faster rcnn
Stars: ✭ 251 (+195.29%)
Mutual labels:  faster-rcnn
FasterRCNN-pytorch
FasterRCNN is implemented in VGG, ResNet and FPN base.
Stars: ✭ 121 (+42.35%)
Mutual labels:  faster-rcnn
Paddledetection
Object Detection toolkit based on PaddlePaddle. It supports object detection, instance segmentation, multiple object tracking and real-time multi-person keypoint detection.
Stars: ✭ 5,799 (+6722.35%)
Mutual labels:  faster-rcnn
Icevision
End-to-End Object Detection Framework - Pluggable to any Training Library: Fastai, Pytorch-Lightning with more to come
Stars: ✭ 218 (+156.47%)
Mutual labels:  faster-rcnn
Shadowless
A Fast and Open Source Autonomous Perception System.
Stars: ✭ 29 (-65.88%)
Mutual labels:  faster-rcnn
Luminoth
Deep Learning toolkit for Computer Vision.
Stars: ✭ 2,386 (+2707.06%)
Mutual labels:  faster-rcnn
GIouloss CIouloss caffe
Caffe version Generalized & Distance & Complete Iou loss Implementation for Faster RCNN/FPN bbox regression
Stars: ✭ 42 (-50.59%)
Mutual labels:  faster-rcnn
Faster Rcnn tf
Faster-RCNN in Tensorflow
Stars: ✭ 2,349 (+2663.53%)
Mutual labels:  faster-rcnn
py-faster-rcnn-imagenet
Train faster rcnn on imagine dataset, related blog post: https://andrewliao11.github.io/object/detection/2016/07/23/detection/
Stars: ✭ 133 (+56.47%)
Mutual labels:  faster-rcnn
Object-Detection-And-Tracking
Target detection in the first frame and Tracking target by SiamRPN.
Stars: ✭ 33 (-61.18%)
Mutual labels:  faster-rcnn
Real-Time-Object-Detection-API-using-TensorFlow
A Transfer Learning based Object Detection API that detects all objects in an image, video or live webcam. An SSD model and a Faster R-CNN model was pretrained on Mobile net coco dataset along with a label map in Tensorflow. This model were used to detect objects captured in an image, video or real time webcam. Open CV was used for streaming obj…
Stars: ✭ 50 (-41.18%)
Mutual labels:  faster-rcnn
MMTOD
Multi-modal Thermal Object Detector
Stars: ✭ 38 (-55.29%)
Mutual labels:  faster-rcnn
View All Similar Projects ➔

keras-faster-rcnn

keras 复现论文 Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks ;主要参考了工程Mask RCNN; 给出了在Pascal VOC目标检测数据集上的训练和测试过程;

  1. 关键点说明
  2. 依赖
  3. 训练
  4. 预测
  5. 评估
  6. 自定义度量
  7. toDoList
  8. 总结

关键点说明:

a.骨干网络使用的是resnet50;Conv14用于提取特征;Conv5用于rcnn分类; 精调Conv3Conv5;所有的Batch Normalization层都不精调，因为batch size太小

b.训练输入图像大小为720*720; 将图像的长边缩放到720,保持长宽比,短边padding;原文是短边600;长边1000

c.不同于源工程，所有的层都是使用的Keras Layer 包括anchors、rpn网络的target、roi2proposals、detect boxes等

d.为了更好的诊断网络增加了10个自定义度量，如：gt数量，rpn网络匹配的gt数，实际训练的正anchor数等

e.anchor尺寸参考yolo使用聚类得到

f. 原文中rpn网络每张图训练256个anchors，正负样本保持1:1；由于PASCAL VOC数据集平均的GT数约为3个，rpn网络实际上能够匹配到的正anchors平均约21个；所以这里rpn网络每张图训练80个anchors

g. 原文中rpn会产生2000个proposals用于rcnn网络训练，但是这里训练几个epoch后，仅仅能够产生不到300个proposals,所有将训练和预测的阶段proposals阶段的iou阈值设置为0.9，原文中为0.7

依赖

tensorflow-gpu 1.14.0

训练

a) 下载工程

git clone https://github.com/yizt/keras-faster-rcnn.git

b) 下载pascal voc数据集,并解压

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar

c) 下载resnet 50预训练模型

wget https://github.com/fchollet/deep-learning-models/releases/download/v0.2/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5

d) 修改工程faster_rcnn/config.py文件中数据集和预训练模型路径

pretrained_weights = '/opt/pretrained_model/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5'
voc_path = '/opt/dataset/VOCdevkit'

e) 聚类生成anchors长宽

python gt_cluster.py --clusters 9

结果如下(每次聚类结果有细微差异):

h:[53.25, 83.97, 149.73, 187.92, 202.07, 270.18, 357.79, 445.08, 488.2] 
w:[40.64, 128.43, 67.18, 304.62, 141.64, 539.33, 188.27, 336.6, 588.34]
ratio:[1.31, 0.65, 2.23, 0.62, 1.43, 0.5, 1.9, 1.32, 0.83]
ious:{1: 0.68557111415695, 2: 0.7254386539924431, 3: 0.692596649433563, 4: 0.6834624200806161, 5: 0.6157859655510957, 6: 0.6595932146657101, 7: 0.6685312631536455, 8: 0.665801489785847, 9: 0.6669742441576416, 10: 0.6121460291384846, 11: 0.6710420495151501, 12: 0.7141121719703986, 13: 0.720584796915154, 14: 0.5639737876945834, 15: 0.7321887018308443, 16: 0.6533920906006669, 17: 0.6594010218140989, 18: 0.5846891288412122, 19: 0.7055301287233066, 20: 0.6066710334483025}
mean iou:0.6643742977734857

并根据聚类结果设置config.py中的，anchors长宽属性

    RPN_ANCHOR_HEIGHTS = [52.42, 85.64, 143.89, 186.92, 208.26, 266.1, 359.72, 446.26, 484.92]
    RPN_ANCHOR_WIDTHS = [40.85, 132.84, 66.24, 294.26, 135.53, 533.3, 190.26, 339.55, 591.88]

f) end2end方式训练网络

python train.py

预测

a) 预训练模型下载

​ PASCAL VOC 2007训练集上训练好的模型下载地址： frcnn-resnet50.035.h5

b) 执行如下命令即可

python inference.py --weight_path /tmp/frcnn-resnet50.035.h5

样例结果如下;更多预测样例见demo_images目录

examples

评估

执行如下命令即可

python evaluate.py --data_set test --weight_path /tmp/frcnn-resnet50.035.h5

输出结果如下：

预测 4952 张图像,耗时：1048.401778936386 秒
ap:{0: 0.0, 1: 0.6768214760548549, 2: 0.6760312377088041, 3: 0.6122568296243334, 4: 0.6367414618873344, 5: 0.6769495102412151, 6: 0.6631589701474099, 7: 0.5463179678366642, 8: 0.3078317216556386, 9: 0.4427729076167924, 10: 0.48129418164837967, 11: 0.5215081042808085, 12: 0.7005935007417052, 13: 0.5352984925242024, 14: 0.3158921077398846, 15: 0.7757831017666008, 16: 0.6199472930246124, 17: 0.32039103004250324, 18: 0.3694302780767011, 19: 0.659186483792924, 20: 0.6523787880767623}
mAP:0.5595292722244065
整个评估过程耗时：1052.8664383888245 秒

自定义度量

  1. gt_num：图像的GT 个数
  2. positive_anchor_num：rpn网络实际训练正样本anchor数
  3. negative_anchor_num：rpn网络实际训练负样本anchor数
  4. rpn_miss_gt_num：rpn网络没有匹配anchor的GT数
  5. rpn_gt_min_max_iou：设max_iou为每个GT匹配的anchor最大IoU;rpn_gt_min_max_iou是每个mini-batch中max_iou的最小值;用于衡量anchor长宽尺寸和个数设置是否合理
  6. roi_num：经过proposal层nms后实际喂入rcnn网络的proposal个数
  7. positive_roi_num：每张图像rcnn网络实际训练的正样本数
  8. rcnn_miss_gt_num：rcnn网络按照0.5的iou阈值匹配，有多少个GT没有匹配到proposals；
  9. rcnn_miss_gt_num_as：在经过正样本比例(0.25)限制后, 有多少个GT没有匹配到proposals；
  10. gt_min_max_iou: 设max_iou为每个GT匹配的proposals最大IoU;gt_min_max_iou是每个mini-batch中max_iou的最小值;用于衡量rpn网络对于边框位置的改善程度

toDoList

  1. 评估标准(已完成)
  2. 边框裁剪放到Anchors层外面(已完成)
  3. batch_slice部分重构(已完成)
  4. 模型编译重构,分离度量指标(已完成)
  5. GT boxes信息加载部分重构(已完成)
  6. 不同输入尺寸大小精度比较
  7. indices类型改为tf.int64;float32类型转换可能丢失精度(已完成)
  8. 使用聚类GT boxes设计anchors长宽尺寸(已完成)

总结

  1. 精度还是与原文相差较大，目前在VOC 2007测试集上的mAP为55.9%；过拟合(在训练集上测试78%+)还是BN层没有训练(batch size太小，训练后效果更差)?
Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].