Registration
- Registration (The contest is open to both industry and academia.)
Target Platform
The 2023 contest will use the Kria KV260 Vision AI Starter Kit and the Jetson Nano Developer Kit.
GPU Platform
The base design framework is provided here: https://github.com/dac-sdc/gpu_starter_2023. This repository is a work in progress and should be available by late March
FPGA Platform
You should use the Ubuntu 22.04 PYNQ PYNQ image, available at http://www.pynq.io/board.html.
The base design framework, with additional setup instructions, is provided here: https://github.com/dac-sdc/fpga_starter_2023.
Training Dataset
Link to download training dataset: https://drive.google.com/file/d/1ceQ5y_rCReSZ26HzzCf2muDNbovjyl5k/view?usp=share_link
The training dataset contains:
- 10000 training images in the JPEGImages directory:
- Some extra images with _1 suffix are included such as 00009_1.jpg, which are simply mirror images of the base image and can be ignored.
- The images have varied size, with the following distribution of sizes:
- 1920x1080: 4962 images
- 1280x720: 4552 images
- 1920x1088: 404 images
- 1920x1072: 82 images
- Labelled object types and locations in the label directory. For example, JPEGImages/00001.jpg will have an associated label/00001.json file describe object types and locations in the image.
- Object types:
| Type | Name | Example | Identifying Data |
|---|---|---|---|
| 1 | Motor Vehicle |  |
Bounding Box |
| 2 | Non-motorized Vehicle |  |
Bounding Box |
| 3 | Pedestrian |  |
Bounding Box |
| 4 | Red Traffic Light |  |
Bounding Box |
| 5 | Yellow Traffic Light |  |
Bounding Box |
| 6 | Green Traffic Light |  |
Bounding Box |
| 7 | Off Traffic Light |  |
Bounding Box |
- Segmentation types
| Type | Name | Example | Identifying Data |
|---|---|---|---|
| 8 | Solid lane line |  |
segment connected component |
| 9 | Dotted lane line |  |
segment connected component |
| 10 | Crosswalk |  |
segment connected component |
- Label descriptions
| Attribute | Field type | Meaning | Example |
|---|---|---|---|
| type | Int | type number (value range 1-10) | 1 |
| x | Float | Abscissa of upper left corner of bounding box | 100.00 |
| y | Float | Ordinate of upper left corner of bounding box | 110.00 |
| width | Float | Width of bounding box | 40.00 |
| height | Float | Height of bounding box | 60.00 |
| segmentation | [polygon] | Segment type connected component. Only type 8-10 have values. Two consecutive numbers represent the coordinate of a point | [[100.00, 110.00, 101.00,111.00…]] |
Test Dataset
The hidden test dataset contains 6000 images. The images have varied size, with the following distribution of sizes:
- 1920x1080: 3788 images
- 1280x720: 1932 images
- 1920x1088: 243 images
- 1920x1072: 37 images
Note we will only use 3000 images for the preliminary submissions and use all 6000 images for the final evaluation. All the designs (including latency and accuracy) will be evaluated based on Batch Size = 1.
Other Info
Frequently Asked Questions
Previous Contest Winning Designs
Note: These are designs for the GPU and FPGA contest, and were for a different image detection problem/dataset than this year.
- 2023: https://github.com/PKU-SEC-Lab/dac-sdc-2023-designs
- 2022: https://github.com/jgoeders/dac_sdc_2022_designs
- 2021: https://github.com/jgoeders/dac_sdc_2021_designs
- 2020: https://github.com/jgoeders/dac_sdc_2020_designs
- 2019: https://github.com/xyzxinyizhang/2019-DAC-System-Design-Contest
- 2018: https://github.com/xyzxinyizhang/2018-DAC-System-Design-Contest
Q&A Platform
See Slack link on sidebar.