Dmytro Khalii
Offer Dmytro work on your next project.
Private message
Offer work
Kharkiv,
Ukraine
2 months 23 days back
available for hire
on the service 4 months 5 days
Rating
Successful projects
Average rating
Skills and abilities
Programming
Services
Photo, Audio & Video
Portfolio
-
Annotation of train details
AI & Machine Learning
Annotation of various train details as a Bitmap object. Photo project.
-
Annotation of transport
AI & Machine Learning
From the project "Ego-view (driver's perspective). Annotation for autonomous vehicles."
Here I teach AI to think like a driver, predicting possibilities and understanding depth. Here is the full technical description of this project:
1. Perspective: the view from the Ego-vehicle's side
… In AI, the Ego-vehicle is a car equipped with cameras. You annotate everything "through the eyes" of the autonomous vehicle. This view is critically important, as the AI needs to calculate the distance from itself to every other object.
2. Occlusion processing ("hidden" parts)
This is the part where you annotate cars as "blocked by obstacles." There are two specific points:
* Modal annotation: I draw a box only around the visible part of the car (for example, only the rear lights sticking out from behind a truck).
* Amodal annotation: This is the "logical" part. I draw a box where the entire car would be, even those parts that I cannot see. This way, I have to "guess" the full size of the car based on its type.
3. Reachability and path logic ("Can I get there?")
I choose which cars will be on the road, which will be blocked by obstacles, and which will be on the shoulder, but the driver's car will be able to approach them. Thus, if there is a car on the side road with no physical barrier (such as a curb or wall) between it and the main road, you annotate that area as "reachable." This teaches the AI that it can turn there to avoid an accident or follow the navigation route.
4. "Trimming" and its degree of annotated cars by any obstacles.
a) When a car is "trimmed" by the edge of the screen.
b) When another object (tree, trash can, pedestrian, or another car) is between the "driver's seat" and the car that I am annotating.