#help choosing a dataset for my Car recognition project via video using python tensorflow,opencv

40 messages · Page 1 of 1 (latest)

void valve
#

I want to start a project using python,opencv and tensorflow to create a car recognition app that will detect other cars from a video camera that's placed in your own car. My question is first, do you guys know any good car datasets? second, do I need to look for a dataset that has pictures/videos/labels/any feature that is filmed from a car perspective or is it enough to find many images of cars from any angle.

I tried looking for many datasets but I couldn't find many that are filmed from a car perspective and many of them were hard to fit in my tensorflow model.

I am new to tensorflow so I don't know if these are the types of questions you ask here but I am trying my best to describe my problem since I am yet to understand much about machine learning and all. thanks for any helpers!

final lotus
pine jungle
#

Do u wanna recognise car model/make?

#

Or just cars?

void valve
#

just cars

pine jungle
#

Oh

#

Cityscape is a fairly large standard and good dataset

void valve
#

checking it out thanks

#

They need to approve my register Will check back in a few days might need help making a model of it if possible thank you again!

hazy trout
#

@void valve The one and only COCO dataset has TONS of car images for object detection too

#

it is a benchmark dataset used for advancing object detection models. I think you'll find 50k+ car images in there

void valve
#

I will try that too thanks!

pine jungle
#

theres stanford cars

#

but its not car detection, its car classification

#

u also have KITTI vehicle/car dataset

void valve
#

I will try alll of them know thank you Will keep you posted

void valve
hazy trout
#

haven't actually downloaded it myself but it should be simple enough using their API or provided download links

void valve
#

I downloaded the api through github

image.png
#

keep in mind Im new to all these stuff so I will probably ask dumb questions

#

now second, I don't know how to activate the api or exactly what it means but then which one of those should I download

image.png
#

@hazy trout sorry for interrupting with stupid questions

hazy trout
#

then its corresponding annotations to the right

void valve
#

how did you know which one to download?

hazy trout
#

the train/val/test are subsets of the full dataset, so you can choose whichever you want, and the 2017 version is the most recent COCO dataset released and all the object detection benchmarks are conducted on that one

#

plus, it's the largest of them all

#

so you would get the most car images

void valve
#

I understand

hazy trout
#

keep in mind though, there are 80 objects labelled in COCO

#

so you need to filter out the images that you need based on the annotations

#

for example, delete all the images that don't have a "car" label in them

void valve
#

ok thank you Will do

void valve
#

How can I run the make file on windows?
Do I must use the api or I can manage without it?

hazy trout
#

you can without it i think, again i didn't download it before so i don't know the exact steps

#

as for the makefile, you can't. that's unix-only

void valve
#

ok sure, I found out the category id for cars so do I delete every image thats not that category? and in the jsons aswell? that seem like a lot of work

rapid herald
#

You may want to try out the open-images-v7 dataset and explore on cars. Follow the fiftyone tools’ documentation for quick filtering