All Projects → beader → tianchi-fashionai

beader / tianchi-fashionai

Licence: other
FashionAI全球挑战赛——服饰关键点定位

Programming Languages

Jupyter Notebook
11667 projects
python
139335 projects - #7 most used programming language

Projects that are alternatives of or similar to tianchi-fashionai

TianChiProj
🏆🥈🥉 TianChiProj 天池
Stars: ✭ 22 (+4.76%)
Mutual labels:  tianchi
rl-medical
Communicative Multiagent Deep Reinforcement Learning for Anatomical Landmark Detection using PyTorch.
Stars: ✭ 36 (+71.43%)
Mutual labels:  landmark-detection
Make Sense
Free to use online tool for labelling photos. https://makesense.ai
Stars: ✭ 2,087 (+9838.1%)
Mutual labels:  landmark-detection
tianchi-3rd security
第三届阿里云安全算法挑战赛
Stars: ✭ 35 (+66.67%)
Mutual labels:  tianchi
Deep-MVLM
A tool for precisely placing 3D landmarks on 3D facial scans based on the paper "Multi-view Consensus CNN for 3D Facial Landmark Placement"
Stars: ✭ 71 (+238.1%)
Mutual labels:  landmark-detection
google-cloud-vision-php
A simple php wrapper for the google cloud vision API
Stars: ✭ 16 (-23.81%)
Mutual labels:  landmark-detection
TC-YoukuVSRE
天池2019阿里巴巴优酷视频增强和超分辨率挑战赛自用代码,EDVR、WDSR、ESRGAN三个模型。
Stars: ✭ 41 (+95.24%)
Mutual labels:  tianchi
visual-compatibility
Context-Aware Visual Compatibility Prediction (https://arxiv.org/abs/1902.03646)
Stars: ✭ 92 (+338.1%)
Mutual labels:  fashionai
dlib-minified-models
A set of minified (but still accurate) models for Dlib
Stars: ✭ 61 (+190.48%)
Mutual labels:  landmark-detection
Mobilenet Yolo
MobileNetV2-YoloV3-Nano: 0.5BFlops 3MB HUAWEI P40: 6ms/img, YoloFace-500k:0.1Bflops 420KB🔥🔥🔥
Stars: ✭ 1,566 (+7357.14%)
Mutual labels:  landmark-detection
deep alignment network pytorch
PyTorch Implementation of the Deep Alignment Network
Stars: ✭ 37 (+76.19%)
Mutual labels:  landmark-detection
latent-pose-reenactment
The authors' implementation of the "Neural Head Reenactment with Latent Pose Descriptors" (CVPR 2020) paper.
Stars: ✭ 132 (+528.57%)
Mutual labels:  landmark-detection
tf-blazepose
BlazePose - Super fast human pose detection on Tensorflow 2.x
Stars: ✭ 139 (+561.9%)
Mutual labels:  landmark-detection
tianchi-diabetes
天池精准医疗大赛——人工智能辅助糖尿病遗传风险预测 第一赛季
Stars: ✭ 20 (-4.76%)
Mutual labels:  tianchi
bob
Bob is a free signal-processing and machine learning toolbox originally developed by the Biometrics group at Idiap Research Institute, in Switzerland. - Mirrored from https://gitlab.idiap.ch/bob/bob
Stars: ✭ 38 (+80.95%)
Mutual labels:  landmark-detection
The-Purchase-and-Redemption-Forecast-Challenge-baseline
天池“资金流入流出预测——挑战baseline”的解决方案,线上效果143.5
Stars: ✭ 78 (+271.43%)
Mutual labels:  tianchi
DeepVTB
🌌 OpenVTuber-虚拟アイドル共享计划 An application of real-time face and gaze analyzation via deep nerual networks.
Stars: ✭ 32 (+52.38%)
Mutual labels:  landmark-detection
YOLO Universal Anatomical Landmark Detection
[MICCAI 2021] You Only Learn Once: Universal Anatomical Landmark Detection https://arxiv.org/abs/2103.04657
Stars: ✭ 40 (+90.48%)
Mutual labels:  landmark-detection
Fatigue-Detection
Eye state classification using OpenCV and DLib to estimate Percentage Eye Closure (PERCLOS) and alert a drowsy person (such as a driver).
Stars: ✭ 46 (+119.05%)
Mutual labels:  landmark-detection
Facial-Landmark-Detection
Facial Landmark Detection Framework
Stars: ✭ 33 (+57.14%)
Mutual labels:  landmark-detection

FashionAI全球挑战赛——服饰关键点定位

竞赛链接

本项目为快速跑通流程的一个原型,可以当成一个 baseline ,线下测试的分数大约为 13% , 最终线上的成绩为 11.81%,在初赛中排 101 名。项目用 Keras 实现。

model

模型部分非常简单,直接使用了 xception net (不使用预训练好的模型),对输出层做了微小改动,使其输出直接是 24个坐标点,共48个数值。

loss

比赛的评分指标用的是 Normalized Error, 其实就是对图片中每个关键点的预测偏差除以了一个归一化参数, 其中上衣外套连衣裙的归一化参数是两个腋窝点的欧式距离,裤子和半身裙的归一化参数是两个裤头的欧式距离。

NE 不直接可导,考虑用 NE 的平方作为 loss。但 NE 平方展开较为复杂,你会发现分子部分实际上有很多二次交叉项, 代码实现起来感觉很繁琐。因此我在实际训练的时候使用的 loss,是直接对 square error 做归一化,再求和做平均。

dataloader

这个步骤比较繁琐,因为数据量还是不小的,无法全部读入内存中,因此把他抽象成一个类来,效果如下:

train_imgs = ImageData(..., source_shape=(512, 512), target_shape=(512, 512)
train_imgs[0]
train_imgs[:100]

train_imgs 像一个 numpy array 一样可以做 index 和 slice,并且会先对图片进行 padding,然后做缩放。

training

训练的时候把原始训练集按照 9: 1 进行切分,对验证集的 loss 进行监控,当 loss 连续 3 个 epochs 不下降的时候,就 调低 learning rate。当验证集的 loss 连续 5 个 epochs 不下降时,就停止训练。每个 epoch 都查看一下验证集 loss 是否 变小,如果变小了就把当前最好的模型保存起来。

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].