Computer Vision talk

Title: "Knowing when you're wrong... and by how much" Duration: Approx. 30 min

---

The talk is about my bachelor thesis: Using physical sensors (TOF / RGB cameras) and monocular depth estimation models (models from now on), together with product of gaussian PDFs to produce better depth estimations.

Physical sensors suffer from noise and the real world being messy, models suffer from not interacting with the real world ("imagining" plausible values). The goal is to "ground" models.

Why gaussians? central limit theorem.

We used synthetic data (photorealistic renders where depth is perfectly known), so we could take an RGB image, an estimated depth, and train a variance estimation model.

Then:

1. Capture RGBD image
2. Split RGBD into RGB and D_sensor
3. Pass RGB through depth estimation model, obtain D_model
4. est_var_sens = F(RGB, D_sensor)
5. est_var_model = F(RGB, D_model)
6. N(mu=D_sensor, var=est_var_sens), N(mu=D_model, var=est_var_model)
7. Product of those gives us N(mu=refined_D, var=refined_var)

Thats the core idea.

Other topics we want to cover:

• Basic error theory
• Model architecture
• Future improvements

Tone

"For all publics": mostly people interested in AI, CV and some maths. It should be informative, and try to not leave anything important out, but not hardcore and "down to the float". It should be understood for people that have a general interest in the field, but are not experts.