Introduction
Depth estimation is traditionally done using a pair of cameras called a stereo pair.
Depth Estimation algorithms using Neural Networks have made enormous strides recently. Some algorithms estimate depth using a single image (monocular depth), others use neural networks to improve the depth estimated using a stereo pair, and some improve the depth estimation for RGBD cameras.
This track aims to develop a high-quality depth estimation project using OpenCV AI Kit with Depth Pro (OAK-D-Pro) Spatial AI Camera. This camera was designed and developed by our partner Luxonis.
OAK-D-Pro uses the stereo pair for depth estimation, although independent researchers have also used all three cameras to improve results. The goal is to use neural networks to improve depth estimation on this device by combining traditional computer vision techniques and modern Deep Learning approaches. You can use one, two, or all three cameras on OAK-D-Pro.
OAK-D-Pro uses the well-documented Python API called Depth-AI API. Participants can also ask questions on the Luxonis Discord Channel.
OAK-D-Pro consists of a stereo pair and an RGB camera. In addition, it has an Intel Myriad X processor capable of 4 TOPS processing power (1.4 TOPS for AI). OAK-D-Pro can be thought of as an RGBD camera with a powerful neural processor that can be used to improve depth estimates.
Previous
Next
Two-Stage Track
This competition track will consist of two stages - Prelims & Finals.
Prelims
Participants can enter the Prelims Stage of the competition by submitting a proposal with the following details by 23:59:59 September 19th 2022 UTC-8
1. Team
- Team Name: Make it fun.
- Team members: Names and brief bios of up to 4 team members. Showing your work with a link to GitHub is also recommended.
- Details: Provide details of what algorithms you are planning you use. You may submit multiple ideas that you are planning to try.
- Background: Demonstrating good knowledge of existing research in the area will improve your chances. So provide copious references.
Finals
Twenty-five teams will be selected for the Finals based on the strength of your proposal. Each team will receive one OAK-D-Pro within 15 days of announcement.
We will release a list of scenarios on which we will test the algorithm.
The teams must submit their final results by 23:59:59 December 18th 2022 UTC-8. The final submission will consist of -
1. Working code: 80% of the credit will be awarded for the working code and/or model with instructions on how to run it on an OAK-D-Pro.
- Code running at 15 fps or faster will be judged only on quality.
- There will be a penalty for code running slower than 15 fps. This means that slower code needs to produce very good results to win.
Prizes
Prelims
The top 25 teams selected for
the finals will receive an OAK-D-Pro.
Finals
The top three winning teams
will receive $5k, $3k, and $2k
cash awards, respectively.
OAK-D-Pro Specs
This track is based around improving depth estimation for OAK-D-Pro.
- Active Laser Depth: depth with no ambient light
- Active Blanket IR illumination: computer vision with no ambient light
- 4 TOPS of processing power (1.4 TOPS for AI)
- Run any AI model, even custom architectured/built onesRun any AI model, even custom architectured/built ones
- Encoding: H.264, H.265, MJPEG - 4K/30FPS, 1080P/60FPS
- Computer vision: warp/dewarp, resize, crop, edge detection, feature tracking
- Stereo depth perception at 120 FPS with filtering
- Dimension: 97x29.5x22.9 mm
- Weight: 91g
Relevant Specifications
Feature |
Color camera |
Stereo pair |
|---|---|---|
|
Sensor
|
IMX378
|
OV9282
|
|
DFOV / HFOV / VFOV
|
81° / 69° / 55°
|
81° / 72° / 49°
|
|
Resolution
|
12MP (4032x3040)
|
1MP (1280x800)
|
|
Focus
|
Auto-Focus: 8cm - ∞
|
Fixed-Focus: 19.6cm - ∞
|
|
Max Framerate
|
60 FPS
|
120 FPS
|
|
F-number
|
2.0
|
2.2
|
|
Lens size
|
1/2.3 inch
|
1/4 inch
|
Laser dot projector projects many small dots in front of the device, which helps with disparity matching, especially for low-visual-interest surfaces (blank surfaces with little to no texture), such as a wall or floor. The technique we use is called ASV – conventional active stereo vision – as stereo matching is performed on the device the same way as on OAK-D (passive stereo).
