50+ Object Detection Datasets from different industry domains

Last Updated on October 10, 2020 by Editorial Team

Author(s): Abhishek Annamraju

Computer Vision

A list of object detection and image segmentation datasets (With colab notebooks for training and inference) to explore and experiment with different algorithms on!

Computer Vision is such a fast-paced field that everyday loads of new techniques and algorithms are presented in different conferences and journals. When it comes to object detection, theoretically you learn about multitudes of algorithms like Faster-rcnn, Mask-rcnn, Yolo, SSD, Retinenet, Cascaded-rcnn, Peleenet, EfficientDet, CornerNet…. This list is never-ending!

It is always beneficial to consolidate your learning experience by applying it on different datasets!!!!

This way you tend to understand the algorithms better, plus you get an intuition over which algorithms work on what kind of datasets.

Our opensource team at Monk Computer Vision Org compiled a list of object detection, image segmentation and action recognition datasets and created short tutorials over each of them for you to utilize these datasets and try out different object detection algorithms

Mentioned below is a shortlist of object detection datasets, brief details on the same, and steps to utilize them. The datasets are from the following domains

★ Agriculture
★ Advance Driver Assistance and Self Driving Car Systems
★ Fashion, Retail, and Marketing
★ Wildlife
★ Sports
★ Satellite Imaging
★ Medical Imaging
★ Security and Surveillance
★ Underwater Imaging

….. and much more!!!!!

The complete list at one place is available with associated usage instructions and training codes on github

Agriculture-related datasets

A) Winegrape Detection Dataset

* Goal — To detect grape clusters in vineyards
* Application — To monitor growth and analyze yield
* Details — 300 images with 4400 bounding boxes over 5 classes of grapes
* How to utilize the dataset and build a custom detector using YoloV3 pipeline

B) Global Wheat Detection Dataset

* Goal — To detect wheat crop from in fields
* Application — To monitor growth and analyze yield
* Details —3430 images with 100K+ annotations
* How to utilize the dataset and build a custom detector using EfficientDet-D4 pipeline

Advance Driver Assistance and Self Driving Car Systems related Datasets

A) LISA Traffic Sign Detection Dataset

* Goal — To detect and classify traffic signs in dash cam images
* Application — Traffic sign recognition acts as a rule setter for autonomous driving
* Details — 7855 annotations on 6610 frames over 47 US sign types
* How to utilize the dataset and build a custom detector using EfficientDet-D3 pipeline

* This repository has one more dataset
 — LISA Vehicle Detection Data

B) Object Detection in Low Lighting Conditions

* Goal — To detect on-road objects in low-lighting conditions — fog, dark, haze, rains, etc
* Application — This is a crucial component in self-driving vehicles as it pertains to a safer vehicle if it’s capable of detecting objects in adverse conditions
* Details —15K+ annotations on 7500 frames over 12 different object types
* How to utilize the dataset and build a custom detector using EfficientDet-D3 pipeline

C) LARA Traffic Lights Detection Dataset

* Goal — To detect traffic lights and classify them as red, green, and yellow
* Application — This does rule-setting for adas and self-driving car systems at road network junctions
* Details — 11K frames with 20K+ annotations over three classes of traffic lights
* How to utilize the dataset and build a custom detector using Mmdet-Faster-Rcnn-fpn50 pipeline

D) Person Detection using Infrared Images

* Goal — To detect people in infrared imagery
* Application — Autonomous vehicles are equipped with infrared cams to detect objects in adverse conditions
* Details — 30 video sequences with 1K+ annotations
* How to utilize the dataset and build a custom detector using Mx-Rcnn pipeline

E) Pothole Detection Dataset

* Goal — To detect potholes from on-road imagery
* Application — Detecting road terrain and potholes results in smooth driving.
* Details — 700 images with 3K+ annotations on potholes
* How to utilize the dataset and build a custom detector using M-Rcnn pipeline

F) Nexet Vehicle Detection Dataset

* Goal — To detect vehicles on-road imagery
* Application — Detecting vehicles is a prime component in autonomous driving
* Details — 7000 images with 15K+ annotations on 6 types of vehicles
* How to utilize the dataset and build a custom detector using Tensorflow Object Detection API

G) BDD100K Adas Dataset

* Goal — To detect on-road objects
* Application — Detecting vehicles, traffic signs, and people is a prime component in autonomous driving
* Details —100K images with 250K+ annotations on 10 types of objects
* How to utilize the dataset and build a custom detector using Tensorflow Object Detection API

H) Linkopings Traffic Signs Dataset

* Goal — To detect traffic signs in images
* Application — Detecting traffic signs is the first step towards understanding traffic rules
* Details —3K images with 5K+ annotations on 40+ types of traffic signs
* How to utilize the dataset and build a custom detector using Mmdet — Cascade Mask Rcnn

Fashion, Retail, and Marketing related Dataset

A) Billboard Detection (Subsampling OpenImages Dataset) Dataset

* Goal — To detect billboards in images
* Application — Detecting billboards forms a crucial part in auto-analyzing marketing campaigns across the city
* Details — 2K images with 5K+ annotations on billboards
* How to utilize the dataset and build a custom detector using Retinanet

B) DeepFashion2 Fashion element Detection Dataset

* Goal — To detect fashion products, clothing, and accessories in images
* Application — Fashion detection has huge applications from data sorting to recommendation engines
* Details — 490K images with around 100s of annotation objects classes
* How to utilize the dataset and build a custom detector using CornetNet-Lite Pipeline

* Another Fashion related dataset is Taobao Commodity Dataset

C) Qmul-OpenLogo Logo Detection Dataset

* Goal — To detect different logos in natural images
* Application — Analyzing frequency of logo appearance in videos and natural scenes is crucial in marketing
* Details — 16K training images with logos from all kinds of brands — food, vehicles, restaurant-chains, delivery services, airlines, etc
* How to utilize the dataset and build a custom detector using mx-rcnn pipeline

Sports-Related Datasets

A) Football Detection Dataset (Subsampling from OpenImages Dataset)

* Goal — To detect football across frames in videos
* Application — Detecting football positions is crucial in auto-analysing situations such as offsides, etc
* Details — Around 3K training images.
* How to utilize the dataset and build a custom detector using yolo-v3 pipeline

B) Playing Card Type Detection

* Goal — To detect playing card in natural images and classify the card type
* Application — Possible application is in analyzing winning odds in different card games
* Details — 500+ images over 52 card class types
* How to utilize the dataset and build a custom detector using mx-rcnn pipeline

C) Soccer Player Detection in Thermal Imagery

* Goal — To localize and track players using thermal imagery
* Application — Tracking players in the game is a crucial part in generating analytics
* Details —3K+ images over 5K+ annotations.
* How to utilize the dataset and build a custom detector using mmdet faster-rcnn pipeline

Security and Surveillance Related Datasets

A) MIO-TCD Vehicle Detection in CCTV Traffic Cams

* Goal — To detect vehicles in cctv traffic cameras
* Application — Detecting vehicles in cctv traffic cams forms a crucial part in security surveillance applications
* Details — 113K images with 200K+ annotations on 5+ types of vehicles
* How to utilize the dataset and build a custom detector using Mmdet — Retinanet pipeline

B) WIDER Person Detection Dataset

* Goal — To detect people in cctv and natural scene images and videos
* Application — CCTV based people detection forms the core of security and surveillance applications
* Details — 10K+ images with 20K+ annotations on detecting pedestrians
* How to utilize the dataset and build a custom detector using Cornernet-Lite pipeline

C) Protective Gear — Helmet and Vest Detection

* Goal — To detect helmet and vests on people
* Application — This forms an integral part in security compliance monitoring
* Details — 1.5K+ images with 2K+ annotations on detecting people, helmets, and vests
* How to utilize the dataset and build a custom detector using Mmdet — Cascade RPN

D) Anomaly Detection in Videos

* Goal — To classify videos as per actions being carried out in videos
* Application — Detecting anomalies in real time helps in stopping crime
* Details — 1K+ videos corresponding to 10 anomaly classes.
* How to utilize the dataset and build a custom classifier using mmaction-tsn50 pipeline

Medical Imaging Datasets

A) Ultrasound Brachial Plexus (BP) Nerve Segmentation Dataset

* Goal — To segment certain nerve types in ultrasound images
* Application — This helps in improving pain management through the use of indwelling catheters that block or mitigate pain at the source.
* Details — 11K+ images with associated instance masks for detecting nerves
* How to utilize the dataset and build a custom detector

B) PanNuke Cancer Instance Segmentation in Cells

* Goal — To segment different cell types in the slide image
* Application — Auto-analyzing presence of cancerous and dead cells in terabytes of data
* Details — 3K+ images with associated instance masks for detecting different cell types
* How to utilize the dataset and build a custom detector

Satellite Imaging Datasets

A) Road Segmentation in Satellite Imagery

* Goal — To segment road lines in satellite imagery
* Application — Helps in urban planning and monitoring roadways
* Details — 1K+ images with associated instance masks for detecting different road regions
* How to utilize the dataset and build a custom detector

B) Traversable region segmentation in Synthetically generated lunar imagery

* Goal — To segment out rocks and find traversable region in lunar imagery
* Application — Essential element in autonomous rovers’ path planning
* Details — 10K+ images with associated instance masks for detecting different rocks and flat ground
* How to utilize the dataset and build a custom detector

C) Cars and Swimming Pools Detection in Satellite Imagery

* Goal — To detect vehicles and pools in satellite imagery
* Application — This forms a crucial part in property tax estimation
* Details — 3.5K+ images with 5K+ annotations labels on cars and pools
* How to utilize the dataset and build a custom detector using cornernet-lite pipeline

D) Roads and Residential area segmentation in Aerial Imagery

* Goal — To segment road and residential areas in satellite imagery
* Application — This forms a crucial part in property tax estimation
* Details — 100 very high resolution images with segmentation masks
* How to utilize the dataset and build a custom detector

* Another similar road segmentation dataset and associated training code

E) Water Body Segmentation in satellite imagery

* Goal — To segment water bodies in satellite imagery
* Application — Very important to understand how water bodies change and evolve over time
* Details — 100 very high resolution images with segmentation masks
* How to utilize the dataset and build a custom detector

* Another such dataset is DeepGlobe Land Cover Classification and it’s associated usage guidelines

Wildlife Related Datasets

A) Tiger Detection Dataset (Subsampled from OpenImages)

* Goal — To detect tigers in natural and drone images
* Application — To monitor endangered species
* Details — 2K+ images with 4k+ annotations.
* How to utilize the dataset and build a custom detector using cornernet-lite pipeline

* One more such dataset could be Monkey detection dataset and it’s associated tutorial

B) Zebras and Giraffes Detection Dataset

* Goal — To detect zebra and giraffe species in natural and drone images
* Application — To monitor endangered species
* Details — 5K+ images with 5k+ annotations.
* How to utilize the dataset and build a custom detector using efficientdet-d3 pipeline

C) Caltech Cameratrap Dataset

* Goal — To detect animals in trap camera types images
* Application — To monitor endangered species
* Details — 10K+ images with 8k+ annotations.
* How to utilize the dataset and build a custom detector using retinanet pipeline

* One more such cameratrap dataset and associated training code

D) Elephant Detection Dataset (Subsampled from COCO dataset)

* Goal — To detect elephant species in natural and drone images
* Application — To monitor endangered species
* Details — 5K+ images with 5k+ annotations.
* How to utilize the dataset and build a custom detector using mmdet-maskrcnn

Underwater Datasets

A) Detecting Sea Turtles in the wild

* Goal — To detect sea turtles in underwater images
* Application — To monitor endangered species
* Details — 5K+ images with 5k+ annotations.
* How to utilize the dataset and build a custom detector using efficientdet

* A similar dataset to monitor fish species underwater and associated utilization code

B) Underwater trash detection Dataset

* Goal — To detect marine trash
* Application — To monitor and control marine waste issue
* Details — 2K+ images with 5k+ annotations.
* How to utilize the dataset and build a custom detector using efficientdet

* A more complex pixel based trash segmentation dataset and associated codes

C) SUIM underwater object detection dataset

* Goal — To segment underwater objects
* Application — Path planning for autonomous underwater vehicles, track divers and monitor marine species
* Details — 1.5K+ images with 1.5k+ annotation masks.
* How to utilize the dataset and build a custom detector

D) Brackish underwater fish recognition dataset

* Goal — To detect marine species in underwater imagery.
* Application — To monitor marine species
* Details — 89 videos to detect fish, crab, shrimp, jellyfish, starfish
* How to utilize the dataset and build a custom detector using mmdet — faster rcnn pipeline

Text Analysis related datasets

A) Document Layout Detection Dataset

* Goal — To detect document layout for further analysis
* Application — Essential to segment images into different parts so that certain rule based nlp and text recognition can further be applied.
* Details — 5K+ images with 10k+ annotations with labels such as paragraphs, images, headers.
* How to utilize the dataset and build a custom detector using mx-rcnn

* A very similar dataset exists for graphical components detection in documents named IIIT-AR-13K, here’s how to utilize the dataset and train a model on it

B) Total-Text Dataset

* Goal — To localize text in natural scenes
* Application — Essential base component to recognize using OCR
* Details — 1.5K+ images with 5K+ polygonal annotations
* How to utilize the dataset and build a custom detector using Text-Snake pipeline

C) YY-Mnist Simple OCR Dataset

* Goal — To localize number and classify them in white backgroud images
* Application — Essential base component to recognize using OCR
* Details — 1K images with 2K+ annotations over 10 classes
* How to utilize the dataset and build a custom detector using Retinanet pipeline

Other Datasets

A) TACO Trash Detection Dataset

* Goal — To localize and segment all kinds of garbage in images
* Application — Critical component in autonomous bots trying to tackle trash problem in public places
* Details — 10K images with 15K+ annotations over 20+ different classes trash objects
* How to utilize the dataset and build a custom detector using Retinanet pipeline

B) Indoor Scene General Object Detection Dataset

* Goal — To localize and detect indoor objects in images
* Application — Autotag images in real-estate and rental websites with amenities
* Details — 3K+ images with 5K+ annotations over 10+ different classes indoor objects such as electronic-appliances, bed, curtains, chairs, etc
* How to utilize the dataset and build a custom detector using Retinanet pipeline

C) EgoHands Hand Segmentation Dataset

* Goal — To segment hands in natural scenes
* Application — First step towards understanding gestures, with applications in human computer interaction, sign language recognition
* Details — 4.8K+ images with corresponding hand masks.
* How to utilize the dataset and build a custom detector using Retinanet pipeline

D) UCF Action recognition dataset

* Goal — To classify videos as per actions being carried out in videos
* Application — Tagging videos is important in storing and retrieving large number of videos
* Details — 1K+ videos corresponding to 101 action type classes.
* How to utilize the dataset and build a custom classifier using mmaction-tsn50 pipeline

E) Oil Tanks Datasets

* Goal — To detect tanks in satellite imagery
* Application — To keep track of oil tanks
* Details — 10K+ images with 10K+ annotations.
* How to utilize the dataset and build a custom classifier using retinanet pipeline

Other action recognition datasets

A) STAIR Action Recognition dataset and how to train a model on it

B) A2D Action Recognition dataset and how to train a model on it

C) KTH Action Recognition dataset and how to train a model on it

APPENDIX

For more details on the tutorials visit our Github page

Tutorial Credits to all the opensource contributors at the Monk Object Detection Library

---

50+ Object Detection Datasets from different industry domains was originally published in Towards AI — Multidisciplinary Science Journal on Medium, where people are continuing the conversation by highlighting and responding to this story.

Published via Towards AI