VroomID

Car Re-Identification :Vehicle Multiple Identification Project

Comment

What Got Me Started

'if we could teach a computer to recognize each car, no matter how alike they seemed?' This pushed me into this project idea, where I used Siamese networks to create a Car Re-Identification system. Yep, here I turned system into car experts!

What I Learned

the project taught me way more than I expected:

  1. Siamese Networks: are like the matchmakers of the AI world. They help compare two car images and decide if they’re the same or different—focusing on even the tiniest details.

  2. Contours and Edges: Finding the outline of the car? Harder than expected! On bumpy roads, the edges got confused with the surface. My system often thought a random road bump was part of the car.

  3. Lighting Trouble: Bright sunshine or weird lighting would mess with the colors, making it tricky to recognize cars.

How I Built It

implementation steps:

Rough screening:

  • After adding the hidden layer and adding the activation function output, set a threshold so that the output value of each node is 0 or
  • Find the Hamming distance of the node output (the number of different codes between the two codes), and set a threshold so that there is a range that is similar to the detected image. #### Fine filter:
  • Calculate the Euclidean distance on the output of the hidden layer (not binary) and find that there is no distinction between the same cars. #### github code:
  • Retrieve image hidden layer nodes.

dataset

  • Features of the vehicle data set: The vehicle has been cut out and there is little background interference.

Problems I faced:

  • Bumpy roads confused my model When the road was uneven, my model would get confused and think the road was part of the car. Not ideal! I had to do some serious filter-tweaking..

  • Light(My arch-nemesis) If the sun was too bright or the lighting was weird, the car’s color would sometimes change in my model’s eyes. Bright red suddenly didn’t look so red anymore! This was a constant battle.

Conclusion

In the end, I built a system that can spot cars even on uneven road or when the lighting is tricky. It was a fun ride, full of challenges and learning.

detailed : https://github.com/igargikaushik/Car-ReIdentification/blob/main/README.md

Built With

Share this project:

Updates