Goal of the Competition
Overhead satellite imagery provides critical, time-sensitive information for use in arenas such as disaster response, navigation, and security. Most current methods for using aerial imagery assume images are taken from directly overhead, known as near-nadir. However, the first images available are often taken from an angle — they are oblique. Effects from these camera orientations complicate useful tasks such as change detection, vision-aided navigation, and map alignment.
In this challenge, participants made satellite imagery taken from a significant angle more useful for time-sensitive applications such as disaster and emergency response
What's in This Repository
This repository contains code from winning competitors in the Overhead Geopose Challenge.
Winning code for other DrivenData competitions is available in the competition-winners repository.
Winning Submissions
Prediction Contest
All of the models below build on the solution provided in the benchmark blog post: Overhead Geopose Challenge - Benchmark. Additional solution details can be found in the reports folder inside the directory for each submission.
The weights for each winning model can be downloaded from the National Geospatial-Intelligence Agency's (NGA's) DataPort page.
| Place | Team or User | Public Score | Private Score | Summary of Model |
|---|---|---|---|---|
| 1 | selim_sef | 0.902184 | 0.902459 | An EfficientNet V2 L encoder is used instead of the Resnet34 encoder because it has a huge capacity and is less prone to overfitting. The decoder is a UNet with more filters and additional convolution blocks for better handling of fine-grained details. MSE loss would produce imbalance for different cities, depending on building heights. The model is trained with an R2 loss for AGL/MAG outputs, which reflects the final competition metric and is more robust to noisy training data. |
| 2 | bloodaxe | 0.889955 | 0.891393 | I’ve trained a bunch of UNet-like models and averaged their predictions. Sounds simple, yet I used quite heavy encoders (B6 & B7) and custom-made decoders to produce very accurate height map predictions at original resolution. Another crucial part of the solution was extensive custom data augmentation for height, orientation, scale, GSD, and image RGB values. |
| 3 | o__@ | 0.882882 | 0.882801 | I ensembled the VFlow-UNet model using a large input resolution and a large backbone without downsampling. Better results were obtained when the model was trained on all images from the training set. The test set contains images of the same location as the images in the training set. This overlap was identified by image matching to improve the prediction results. |
| 4 | kbrodt | 0.872775 | 0.873057 | The model uses a UNet architecture with various encoders (efficientnet-b{6,7} and senet154) and has only one above-ground level (AGL) head and two heads in the bottleneck for scale and angle. The features are a random 512x512 crop of an aerial image, the city's one hot encoding, and ground sample distance (GSD). The model is trained with mean squared error (MSE) loss function for all targets (AGL, scale, angle) using AdamW optimizer with 1e-4 learning rate. |
Model Write-up Bonus
| Prediction rank | Team or User | Public Score | Private Score | Summary of Model |
|---|---|---|---|---|
| 2 | bloodaxe | 0.889955 | 0.891393 | See the "Prediction Contest" section above |
| 5 | chuchu | 0.856847 | 0.855636 | We conducted an empirical upper bound analysis, which suggested that the main errors are from height prediction and the rest are from angle prediction. To overcome the bottlenecks we proposed HR-VFLOW, which takes HRNet as backbone and adopts simple multi-scale fusion as multi-task decoders to predict height, magnitude, angle, and scale simultaneously. To handle the height variance, we first pretrained the model on all four cities and then transferred the pretrained model to each specific city for better city-wise performance. |
| 7 | vecxoz | 0.852948 | 0.851828 | First, I implemented training with automatic mixed precision in order to speed up training and facilitate experiments with the large architectures. Second, I implemented 7 popular decoder architectures and conducted extensive preliminary research of different combinations of encoders and decoders. For the most promising combinations I ran long training for at least 200 epochs to study best possible scores and training dynamics. Third, I implemented an ensemble using weighted average for height and scale target and circular average for angle target. |
Approved for public release, 21-943
42 Nov 24, 2022
29 Sep 20, 2022
5 Jan 11, 2022
64 Dec 16, 2022
55 Dec 21, 2022
37 Dec 20, 2022
41 Dec 14, 2022
2 Jul 06, 2022
17 Nov 22, 2022
3 Dec 13, 2021
4 Dec 18, 2022
12 May 12, 2022
17 Jan 01, 2023
704 Dec 27, 2022
53 Dec 23, 2022
1.9k Dec 30, 2022
16 Jun 16, 2022
3 Aug 04, 2022
2 Oct 20, 2022
13 May 01, 2022