Beyond the Sparsity Assumption: Arbitrary-Distribution Infrared Small Target Detection

Official implementation of AdaScope — an asynchronous coarse-to-fine framework designed for arbitrary-distribution infrared small target detection (IRSTD). AdaScope decouples target localization from context allocation via a divide-and-conquer strategy, reformulating coarse localization as a spatial existence classification problem and learning distribution-adaptive observation windows through reinforcement learning.

Table of Contents

• Overview
• Installation
• Dataset Preparation
• Training
• Testing
• Results
• Repository Structure

---

Overview

AdaScope is an asynchronous coarse-to-fine detector for arbitrary-distribution infrared small target detection: GEDM performs coarse localization by classifying spatial existence on a semantic grid (instead of fitting unstable cluster boundaries), and ARFR uses reinforcement learning (GRPO) to adaptively refine the observation window of each candidate, with the downstream detector's performance gain as the reward. Global and local predictions are fused by NMS.

Back to top

---

Installation

Step 1: Clone the Repository

git clone git@github.com:GrokCV/AdaScope.git
cd AdaScope

All commands below must be run from the repository root.

Step 2: Create a Conda Environment

conda create --name deepir python=3.9 -y
conda activate deepir

Step 3: Install PyTorch

Install PyTorch with CUDA 11.8 (or a version matching your driver):

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 \
  -c pytorch \
  -c nvidia

Step 4: Install OpenMMLab Codebases

pip install -U openmim

mim install mmengine
mim install "mmcv>=2.0.0"
mim install "mmdet>=3.0.0"

Step 5: Install AdaScope

python setup.py develop
# or
pip install -e .

Back to top

---

Dataset Preparation

AdaScope is developed on the DenseSIRST infrared small-target dataset.

Step 1: Download DenseSIRST

Download the dataset and place it under data/ so that the layout is:

data/DenseSIRST/SIRSTdevkit/
├── PNGImages/                    # infrared images (*.png)
├── Splits/
│   ├── trainval_v2.txt           # training split
│   └── test_v2.txt               # testing split
└── SIRST/
    ├── BBox/                     # single-instance annotations (*.xml)
    └── Cluster_relabel_graph_v2/ # cluster annotations (*_with_clusters.xml)

If your data lives elsewhere, override DATA_ROOT in configs/detection/_base_/datasets/adascope_densesirst.py.

Step 2: Generate Cluster Annotations

AdaScope trains its cluster head and refiner against cluster-level annotations (a box that encloses a group of small targets). Generate them from the single-instance XMLs with the provided tools — run the first, then the second:

# (1) Graph-based clustering of the test split → Cluster_relabel_graph_v1
python tools/relabel_test_clusters_graph.py

# (2) Supplement Cluster_relabel_graph_v1 with singleton clusters so that every
#     instance is covered → Cluster_relabel_graph_v2
python tools/build_cluster2_from_cluster1.py

These two tools reproduce the Cluster_relabel_graph_v2 directory referenced by the training config. See the header of each script for its arguments.

Back to top

---

Training

AdaScope is trained in three stages automatically via the SynWarmupSupGRPOStageHook:

Stage	Epochs	What is trained
Pretrain	1–12	backbone / neck / GEDM cluster head / global head
Center offset SFT	13–16	ARFR (center alignment, supervised)
RL postraining4Scope	17–28	ARFR (context allocation, Scope Adjustment)

Single-GPU Training

CUDA_VISIBLE_DEVICES=0 python tools/train_det.py configs/detection/cluster/AdaScope.py

Multi-GPU Training

CUDA_VISIBLE_DEVICES=0,1,2,3 \
torchrun --nproc_per_node=4 \
tools/train_det.py configs/detection/cluster/AdaScope.py

Checkpoints are saved to work_dirs/adascope_fcos_densesirst/, with the best one tracked by merged_voc/mAP.

RL Policy Variants

The ARFR policy optimizer is pluggable. All variants share the same coarse-to-fine pipeline (GEDM + ARFR + local FCOS) through the common base detector AdaScopeRLDetector — only the detector/refiner classes and the policy-stage name differ:

Config	Policy optimizer
AdaScope.py	GRPO
AdaScope_ppo.py	PPO
AdaScope_sac.py	SAC
AdaScope_trpo.py	TRPO

Train any variant with the same entry point:

CUDA_VISIBLE_DEVICES=0 python tools/train_det.py configs/detection/cluster/AdaScope.py

AdaScope_TOOD.py uses the same AdaScope detector/refiner and cluster head as AdaScope.py, replacing the global FCOS head with a TOOD head. Its default schedule is 20 epochs, matching the TOOD experiments reported in the paper.

Test with the corresponding config and checkpoint:

CUDA_VISIBLE_DEVICES=0 python tools/test_det.py \
  configs/detection/cluster/AdaScope_ppo.py \
  work_dirs/adascope_ppo_densesirst/best_merged_voc_mAP_epoch_XX.pth

Troubleshooting: Empty Cluster GT Data

If the cluster GT XML files are not found at training time, the refiner receives no valid proposals and all policy / refiner losses become zero. The symptom is easy to spot:

• refiner_sup_num_valid stays at 0 (or very low) in the training log
• GPU memory usage is ~650 MiB instead of the normal ~1450 MiB

The base config configs/detection/_base_/datasets/adascope_densesirst.py ships with a placeholder absolute path that does not exist on your machine:

DATA_ROOT = '/path/to/your/DenseSIRST/SIRSTdevkit'

This is intentional — training will fail immediately with a FileNotFoundError instead of silently loading empty cluster GT data. Edit DATA_ROOT to point to your local DenseSIRST directory.

Diagnose

# Check whether cluster GT is loaded (memory should be ~1450 MiB)
grep "memory:" work_dirs/adascope_fcos_densesirst/train.log | head -5

Back to top

---

Testing

CUDA_VISIBLE_DEVICES=0 python tools/test_det.py \
  configs/detection/cluster/AdaScope.py \
  /path/to/best_merged_voc_mAP_epoch_XX.pth

By default the evaluator reports the merged metric (global + local fusion, mAP @ IoU=0.5, 11-point VOC), which is also used for best-checkpoint selection. To additionally report per-branch metrics (global / raw_cluster / refined_cluster / local), set SHOW_ALL_METRICS = True in configs/detection/cluster/AdaScope.py.

Back to top

---

Results

Comparison with state-of-the-art methods on DenseSIRST (mAP and Recall under the VOC 11-point protocol at IoU=0.5; PR / F1 / F2 at a confidence threshold):

Method	Backbone	FLOPs ↓	Params ↓	mAP50 ↑	Recall50 ↑	PR ↑	F1 ↑	F2 ↑
One-stage
FCOS	ResNet50	50.291G	32.113M	0.257	0.315	0.1571	0.2023	0.2445
VFNet	ResNet50	48.317G	32.709M	0.253	0.336	0.1820	0.2251	0.2624
YOLOX	—	8.578G	8.968M	0.210	0.341	0.1450	0.1882	0.2291
TOOD	ResNet50	50.456G	32.018M	0.256	0.355	0.1750	0.2224	0.2655
DyHead	ResNet50	27.866G	38.890M	0.249	0.335	0.1650	0.2090	0.2488
DDOD	ResNet50	46.514G	32.378M	0.253	0.335	0.1620	0.2079	0.2504
GFL	ResNet50	52.296G	32.258M	0.264	0.367	0.1850	0.2331	0.2762
Two-stage
Cascade R-CNN	ResNet50	90.978G	69.152M	0.136	0.188	0.1150	0.1338	0.1484
SABL	ResNet50	0.125T	42.213M	0.124	0.104	0.1050	0.0969	0.0926
Dynamic R-CNN	ResNet50	63.179G	41.348M	0.184	0.235	0.1420	0.1678	0.1883
End2End
Sparse R-CNN	ResNet50	45.274G	0.106G	0.183	0.572	0.0850	0.1444	0.2488
DAB-DETR	ResNet50	28.939G	43.702M	0.005	0.054	0.0100	0.0160	0.0250
DQ-DETR	—	783.57G	58.68M	0.0149	0.154	0.0200	0.0345	0.0610
EFLNet	—	65.49G	38.335M	0.152	0.1349	0.1400	0.1263	0.1193
PConv	—	65.32G	37.136M	0.164	0.179	0.1500	0.1525	0.1540
Coarse2Fine
DMNet	ResNet50	144.45G	63.18M	0.2239	0.1259	0.2400	0.1412	0.1132
AdaZoom	ResNet50	49.96G	31.887M	0.052	0.348	0.0360	0.0554	0.0819
YOLC	HRNet	121.59G	67.55M	0.335	0.477	0.3416	0.3527	0.3597
BAFE-Net	ResNet50	71.639G	35.626M	0.283	0.335	0.4275	0.3302	0.2905
YOLD	HRNet	122.62G	67.61M	0.171	0.371	0.3312	0.2172	0.1800
SSCFNet	—	76.29G	19.24M	0.3086	0.4351	0.9420	0.5704	0.4613
HFMNet	—	11.19G	7.38M	0.4117	0.5096	0.7629	0.6003	0.5322
AdaScope (Ours)	ResNet50	44.95G	30.05M	0.602	0.732	0.6538	0.6291	0.6157

Cross-Dataset

Method	SIRST mAP50	SIRST Recall	SIRST Params(M)	SIRST FLOPs(G)	SIRST FPS	IRSTD-1K mAP50	IRSTD-1K Recall	IRSTD-1K Params(M)	IRSTD-1K FLOPs(G)	IRSTD-1K FPS
AdaScope	82.93	81.65	30.05	44.95	31.87	59.04	63.64	30.05	44.95	31.87
DNANet	78.00	86.24	4.70	57.04	8.35	57.35	74.75	4.70	57.04	8.35
ResUNet	67.12	87.16	0.90	15.32	167.94	55.66	76.77	0.90	15.32	167.94
ISNet	0.00	0.00	1.09	122.47	35.28	0.67	8.75	1.09	122.47	35.28
UIU-Net	70.14	87.16	50.54	217.70	27.59	58.40	75.08	50.54	217.70	27.59
ACM	0.00	0.00	0.40	1.61	170.97	0.00	0.00	0.40	1.61	170.97
ALCNet	9.84	4.59	0.52	1.51	188.51	0.68	8.75	0.52	1.51	188.51
SSCFNet	23.96	80.73	19.24	76.29	13.33	25.35	85.86	19.24	76.29	13.33
HFMNet	39.40	79.91	7.38	11.19	33.25	49.33	86.87	7.38	11.19	33.25

Back to top

---

Repository Structure

AdaScope/
├── configs/
│   └── detection/
│       ├── _base_/
│       │   ├── datasets/adascope_densesirst.py   # dataset + pipelines
│       │   ├── schedules/schedule_1x.py          # LR schedule
│       │   └── default_runtime.py                # runtime defaults
│       └── cluster/
│           ├── AdaScope.py                       # main experiment config (GRPO)
│           ├── AdaScope_ppo.py                   # PPO variant config
│           ├── AdaScope_sac.py                   # SAC variant config
│           ├── AdaScope_trpo.py                  # TRPO variant config
│           └── adascope_fcos_local.py            # external local FCOS
├── deepir/
│   ├── models/
│   │   ├── detectors/
│   │   │   ├── adascope_detector.py              # FixedFlatSyncGRPODetector (GRPO)
│   │   │   ├── adascope_rl_detector.py           # AdaScopeRLDetector (shared base)
│   │   │   ├── adascope_ppo_detector.py          # AdaScopePPODetector
│   │   │   ├── adascope_sac_detector.py          # AdaScopeSACDetector
│   │   │   └── adascope_trpo_detector.py         # AdaScopeTRPODetector
│   │   ├── refine/
│   │   │   ├── adascope_refiner.py               # ARFR (GRPO refiner)
│   │   │   └── adascope_{ppo,sac,trpo}_refiner.py# RL variant refiners
│   │   └── cluster_heads/adascope_cluster_head.py# GEDM (C5 existence classifier)
│   ├── datasets/
│   │   ├── sirst_voc_det.py                      # DenseSIRST dataset
│   │   └── transforms/adascope_cluster_targets.py# cluster GT generation
│   ├── engine/hooks/adascope_stage_hook.py       # shared 3-stage RL training hook
│   └── evaluation/metrics/selective_voc_metric.py# per-branch VOC metrics
├── tools/
│   ├── train_det.py / test_det.py                # training & testing entry
│   ├── relabel_test_clusters_graph.py            # cluster generation (v1)
│   └── build_cluster2_from_cluster1.py           # cluster generation (v2)
└── docs/
    ├── pipeline.png
    └── ARFR.png

Back to top

---

Citation

If you find this work useful, please consider citing:

@article{adascope,
  title={Beyond the Sparsity Assumption: Arbitrary-Distribution Infrared Small Target Detection},
  author={Jingtang Chen, ZhuLiu, Mingjian Fu,Yimian Dai*},
  journal={},
  year={2026}
}

License

This project is released under the Apache 2.0 license.