DaD: Distilled Reinforcement Learning for Diverse Keypoint Detection

Johan Edstedt · Georg Bökman · Mårten Wadenbäck · Michael Felsberg

Paper

DaD's a pretty good keypoint detector, probably the best.

Run

import dad
from PIL import Image
img_path = "assets/0015_A.jpg"
W, H = Image.open(img_path).size# your image shape,
detector = dad.load_DaD()
detections = detector.detect_from_path(
  img_path, 
  num_keypoints = 512,
  return_dense_probs=True)
detections["keypoints"] # 1 x 512 x 2, normalized coordinates of keypoints
detector.to_pixel_coords(detections["keypoints"], H, W)
detections["keypoint_probs"] # 1 x 512, probs of sampled keypoints
detections["dense_probs"] # 1 x H x W, probability map

Visualize

import dad
from dad.utils import visualize_keypoints
detector = dad.load_DaD()
img_path = "assets/0015_A.jpg"
vis_path = "vis/0015_A_dad.jpg"
visualize_keypoints(img_path, vis_path, detector, num_keypoints = 512)

Install

Get uv

curl -LsSf https://astral.sh/uv/install.sh | sh

In an existing env

Assuming you already have some env active:

uv pip install dad@git+https://github.com/Parskatt/dad.git

As a project

For dev, etc:

git clone git@github.com:Parskatt/dad.git
uv sync
source .venv/bin/activate

Evaluation

For to evaluate, e.g., DaD on ScanNet1500 with 512 keypoints, run

python experiments/benchmark.py --detector DaD --num_keypoints 512 --benchmark ScanNet1500

Note: leaving out num_keypoints will run the benchmark for all numbers of keypoints, i.e., [512, 1024, 2048, 4096, 8192].

Third party detectors

We provide wrappers for a somewhat large set of previous detectors,

python experiments/benchmark.py --help

Training

To train our final model from the emergent light and dark detector, run

python experiments/repro_paper_results/distill.py

The emergent models come from running

python experiments/repro_paper_results/rl.py

Note however that the types of detectors that come from this type of training is stochastic, and you may need to do several runs to get a detector that matches our results.

Changes versus paper version

1. I've updated inference to not include border pixels, see https://github.com/Parskatt/dad/issues/3 for details. However, training is still including borders. Results on most benchmarks is similar.

How I run experiments

(Note: You don't have to do this, it's just how I do it.) At the start of a new day I typically run

python new_day.py

This creates a new folder in experiments, e.g., experiments/w11/monday. I then typically just copy the contents of a previous experiment, e.g.,

cp experiments/repro_paper_results/rl.py experiments/w11/monday/new-cool-hparams.py

Change whatever you want to change in experiments/w11/monday/new-cool-hparams.py.

Then run it with

python experiments/w11/monday/new-cool-hparams.py

This will be tracked in wandb as w11-monday-new-cool-hparams in the DaD project.

You might not want to track stuff, and perhaps display some debugstuff, then you can run instead as, which also won't log to wandb

DEBUG=1 python experiments/w11/monday/new-cool-hparams.py

Evaluation Results

MegaDepth1500 Essential

Method ↓	Max № Keypoints
	512	1024	2048	4096	8192
DaD (Paper)	64.9	67.9	69.4	69.3	68.8
DaD (This Repo)	64.8	68.0	69.4	69.6	68.8

MegaDepth1500 Fundamental

Method ↓	Max № Keypoints
	512	1024	2048	4096	8192
DaD (Paper)	50.6	55.5	57.4	57.8	56.4
DaD (This Repo)	50.6	55.5	57.7	58.2	56.4

ScanNet1500 Essential

Method ↓	Max № Keypoints
	512	1024	2048	4096	8192
DaD (Paper)	25.7	27.9	28.3	28.9	29.3
DaD (This Repo)	26.3	28.0	28.6	29.0	29.2

ScanNet1500 Fundamental

Method ↓	Max № Keypoints
	512	1024	2048	4096	8192
DaD (Paper)	18.3	20.6	21.7	22.2	23.0
DaD (This Repo)	18.6	21.4	21.9	22.5	23.3

HPatches Viewpoint

Method ↓	Max № Keypoints
	512	1024	2048	4096	8192
DaD (Paper)	58.2	60.5	61.0	61.5	61.4
DaD (This Repo)	58.3	60.4	61.0	61.2	61.5

Licenses

DaD is MIT licensed.

Third party detectors in dad/detectors/third_party have their own licenses. If you use them, please refer to their respective licenses in here (NOTE: There may be more licenses you need to care about than the ones listed. Before using any third party code, make sure you're following their respective license).

BibTeX

@article{edstedt2025dad,
  title={{DaD: Distilled Reinforcement Learning for Diverse Keypoint Detection}},
  author={Edstedt, Johan and B{\"o}kman, Georg and Wadenb{\"a}ck, M{\aa}rten and Felsberg, Michael},
  journal={arXiv preprint arXiv:2503.07347},
  year={2025}
}