This repository provides a training pipeline for Ultralytics object detection models that replaces the standard label/anchor assignment with the Lightweight Anchor Dynamic Assignment (LADA) algorithm.
LADA is a lightweight, dynamic anchor assignment strategy for selecting positive samples during training. It reduces label noise, improves stability in crowded scenes, and works as anchor-free detection head.
- Dynamic positive/negative selection to reduce label noise.
- Improved training stability in crowded and multi-object scenes.
- Simple to use Integrated into Ultralytics Object Detection Models.
To more LADA Paper
Anchors (or more generally, candidate boxes/points) are a way to propose multiple potential object locations per image region.
This concept originated from Region Proposal Networks (RPN) in Faster R-CNN, and later adapted to one-stage detectors like SSD.
Modern YOLO models:
- Historically used anchor-based heads.
- Newer variants are often anchor-free (predicting distances from grid points, not fixed anchor boxes).
LADA operates at the assignment stage, deciding which candidates become positives for each ground truth (GT) box during training.
In Ultralytics models, the typical architecture is:
Backbone → Head (FPN included) → Detect
Goal: Extract multi-scale semantic features from the image.
Example outputs for strides {8, 16, 32}:
- P3:
[B, C3, H/8, W/8] - P4:
[B, C4, H/16, W/16] - P5:
[B, C5, H/32, W/32]
Goal: Process backbone features to predict bounding box offsets and class probabilities.
Example outputs:
- f3:
[B, 4*regmax + nc, H/8, W/8] - f4:
[B, 4*regmax + nc, H/16, W/16] - f5:
[B, 4*regmax + nc, H/32, W/32]
Goal: Convert regression distributions to usable bounding boxes.
With Distributed Focal Loss (DFL), each box side (L, T, R, B) is predicted as a discrete probability distribution over regmax bins.
- Input: f3, f4, f5 (from Head)
- Output:
[B, 4 + nc, Total_points]in XYXY pixel format
During training, we must assign positive and negative points to compute loss.
The image above shows anchor points (squares centers) for all strides (32, 16, 8, (imgsz=640)). Positive points are anchors which include any part of dog. But all the positive points might not include meaningful features for dog (might be tree).
LADA improves this assignment process using a multi-stage selection strategy:
LADA selection flow from initial candidates to final positives.[LADAModel.train(debug=true]
For each GT box (cx, cy, w, h) and stride s:
r(s) = s / max_stride
EPCP(s, GTBox) = GTBox_width_height * r(s)
- Select all grid points inside EPCP as initial candidates.
For each candidate:
- Decode predictions to box & class score.
- Compute CLA:
Deviation = |l−r|/(l+r) + |t−b|/(t+b)
DeviationLoss = 0 if 0 ≤ dev ≤ 1 else dev − 1
CLA = ClassLoss + GIoULoss + DeviationLoss
- Keep top 9 candidates per stride.
- From all Candidates_v2, keep top 20 overall.
- Compute average CLA of Candidates_v3.
- Keep only candidates above average score.
git clone https://github.com/mevlt01001/LADA-Detection.git
cd LADA-Detection
pip install -r requirements.txtfrom LADAModel import LADAModel
from ultralytics import YOLO, RTDETR
yolo11n = YOLO("yolov11n.pt")
rtdetr_l = RTDETR("rtdetr_l.pt")
models = [yolo11n] # You can fuse models like => models = [yolo11n, rtdetr_l]
model = LADAModel(
models=models,
nc=20, # VOC2012 includes 20 classes
imgsz=640,
device="cuda"
)
model.train(data_yaml="/home/neuron/Downloads/VOC2012/data.yaml", epoch=100, batch=24, debug=True) # debug parameter, used to visualize LADA assignment.