NestDetect

NestDetect is a research implementation of class-incremental object detection with YOLO11n and YOLO11n-HoPe. It studies catastrophic forgetting when a detector learns laptop and book after first learning person, chair, and dining table.

The repository includes reproducible dataset builders, replay and no-replay training protocols, replay-free Continuum Memory System (CMS) ablations, evaluation utilities, five documented checkpoints, a command-line interface, and a Streamlit application.

Research highlights

Forgetting is defined as:

old-class mAP50-95 before incremental training
- old-class mAP50-95 after incremental training

Lower is better. A negative value indicates positive backward transfer.

YOLO11n baseline

Strategy	Old-class mAP50-95	New-class mAP50-95	Forgetting
Base model	0.3778	—	—
Incremental without replay	0.1399	0.5225	0.2379
Incremental with replay	0.3691	0.5265	0.0086

Balanced replay reduced baseline forgetting by approximately 96.4% without reducing new-class performance.

YOLO11n-HoPe and CMS

Strategy	Old-class mAP50-95	New-class mAP50-95	Forgetting
HoPe base model	0.2037	—	—
HoPe without replay	0.0639	0.4349	0.1398
CMS V4 routed memories	0.2037	0.4097	-0.00002
CMS V5 compressed memory	0.2035	0.4099	0.00016
HoPe with replay	0.2572	0.4431	-0.0535
CMS V4 Replay-Fusion	0.2573	0.4487	-0.0536

CMS V5 preserves the HoPe base-task score without storing old-task images. Its plastic parameter delta is compressed by 3.46×. Replay-Fusion has the highest score on the current validation split, but it is replay-informed and was tuned on that split; it is not a replay-free result or a held-out test result.

The complete methodology, interpretation, limitations, and conclusions are in docs/research-report.md.

Published checkpoints

Checkpoint	Intended role	Replay-free
models/nestdetect_hope_cms_only_v4.pt	Default application and prediction-test checkpoint	Yes
models/nestdetect_hope_cms_v5.pt	Compressed replay-free research checkpoint	Yes
models/nestdetect_hope_cms_v4_replay_fusion.pt	Highest current validation score	No
models/nestdetect_hope.pt	Single-detector HoPe replay checkpoint	No
models/nestdetect_final.pt	YOLO11n replay baseline	No

Each checkpoint has a matching metadata file with its SHA-256 digest, provenance, target classes, and evaluation metrics. See docs/model-card.md.

Target classes

The detector preserves the original 80-class COCO head and activates five class IDs during training and inference:

COCO ID	COCO label	Display label
0	person	person
56	chair	chair
60	dining table	table
63	laptop	laptop
73	book	book

Preserving the COCO80 head prevents pretrained classifier tensors from being discarded when the active class set grows.

Repository structure

NestDetect/
├── app.py
├── configs/
│   ├── baseline/
│   ├── hope/
│   ├── ablations/
│   └── models/
├── datasets/                 # manifests and locally generated datasets
├── docs/
├── models/                   # five published checkpoints and metadata
├── results/                  # compact CSV, JSON, and selected figures
├── scripts/                  # dataset, evaluation, plotting, and CMS utilities
├── src/nestdetect/           # installable Python package
├── tests/
└── pyproject.toml

Installation

Requirements:

• Python 3.10 or newer;
• 8 GB RAM or more;
• optional CUDA-compatible GPU for training;
• optional webcam for real-time detection.

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"

On Windows PowerShell, activate the environment with:

.venv\Scripts\Activate.ps1

Run the application

python -m streamlit run app.py

The application uses the replay-free CMS V4 checkpoint for image and webcam prediction tests. The model is selected explicitly, so the interface never changes prediction provenance through an automatic fallback.

Command-line inference

nestdetect detect \
  --model models/nestdetect_hope_cms_only_v4.pt \
  --source sample.jpg \
  --output results/detection_outputs/sample.jpg \
  --conf 0.25 \
  --imgsz 640

Predictions are filtered to the five target classes unless --all-classes is provided.

Tests

pytest

The current suite contains 26 tests covering datasets, replay memory, evaluation, HoPe modules, continual-learning constraints, conditional memory, and optimizers.

Reproduce the experiments

python scripts/prepare_coco_subset.py
python scripts/build_datasets.py

nestdetect train configs/baseline/base.yaml
nestdetect train configs/baseline/incremental-no-replay.yaml
nestdetect train configs/baseline/incremental-replay.yaml

nestdetect train configs/hope/base.yaml
nestdetect train configs/hope/incremental-no-replay.yaml
nestdetect train configs/hope/incremental-replay.yaml

Evaluation and CMS construction commands are documented in docs/reproducibility.md.

Scope and responsible use

The experiments use a small, curated subset of Microsoft COCO 2017 and one deterministic seed. The reported metrics are research evidence for this controlled protocol, not production guarantees. The model should not be the sole basis for decisions affecting safety, rights, or access to services.

The repository does not redistribute COCO images. Dataset details and source terms are documented in docs/dataset.md.

Documentation

• Architecture
• Dataset
• Model card
• Research report
• Replay-free study
• Reproducibility

References

• Nested Learning / HoPe
• Microsoft COCO
• COCO paper
• Ultralytics YOLO documentation
• Streamlit

Citation

If you use this work, please cite:

Athallah, Z. H. (2026). NestDetect: Mengurangi Catastrophic Forgetting pada YOLO melalui Nested Learning dan Continuum Memory System. Zenodo. https://doi.org/10.5281/zenodo.20958461