bounding boxes

[ebgoldstein]

modify the code to be able to export bounding box location + class.. so that this info can be overlain on the image.

[Perri-Woodard]

Hi Evan,

I have updated the code to include exporting a csv with the bounding box location + class, as well added functionality to output images with the bounding boxes drawn and class labels added. Let me know if this seems sufficient/ if I should make any edits! Also, let me know if there is a time next week later in the week when you would be available to discuss next steps. Thanks!

[ebgoldstein]

Awesome! Can you post some results? I recommend taking all the pics you had in the slideshow and posting those with bounding boxes/classes to see what is going on... also- I recommend looking through a bunch of images just to see for yourself if the model is/is not working ...

And Post some pics here in this thread!

[Perri-Woodard]

After looking through some of the images, there are a couple of general takeaways I have:

• Occasional identification of raindrops or water shadows as people occur
• Counts boat in the high ground as a car a couple of times
• Double counts trucks as cars (or motorcycles) if the vehicle is being obstructed by something
• Counts shed in the back as car or truck, or something in the house is being counted as a vehicle
• Will count reflections in water as people, even if it is not a person's reflection
• Will occasionally count people in vehicles
• Counts random objects that are slender in nature as people: equipment, stop signs, poles, fire hydrants, cones, etc.
• Occasionally counts random items off to the side of the images as cars or trucks.

In general, the model works okay. It definitely has some issues with identifying items that are not in the foreground or are off to the side, or double counting parts of one vehicle as two items. I will post some of these examples below!

[Perri-Woodard]

Raindrop as a person:

Boat counted as a car:

My personal favorite, the carck:

The shed (that's apparently a truck):

Pole reflection in right hand corner being counted as a person??:

Person in car:

Random (slender) things as people:

Random items on edges of photos as vehicles:

Overall, my conclusion that person identification definitely needs help (or computers just think people are EXTREMELY skinny.)

[Perri-Woodard]

Here is a link to the updated slideshow with bounding boxes and classes: Object_Detection V1 Results Discussion- with bounding boxes

[ebgoldstein]

awesome - those slides are great @Perri-Woodard .

cc: @anardek - this is an good follow up from last week.

I'm closing this issue, and opening a new one re: labeling data and retraining the model