Research map / Video / Video segmentation and tracking
MOCID: Motion Context and Displacement Information Learning for Moving Infrared Small Target Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MOCID, a spatio-temporal detector for moving infrared small targets that combines Fourier-inspired clip-level motion context (FISTA) guiding dynamic convolutions with a Mamba-based displacement module (DAM) for frame-to-frame displacement.
Problem
Moving infrared small target detection (MIRSTD) struggles because sequential spatial-then-temporal modules do not inject motion/displacement differences between targets and background noise into spatial feature extraction, causing noise-interfered, error-prone localization.
Approach
A FISTA block uses spatial and temporal DFT with learnable global filters to derive clip-level motion context that calibrates per-frame dynamic convolution weights; a DAM module uses temporal interpolation of target and reference frames plus 3D central difference convolution to parameterize a Mamba selective scan that models frame-to-frame displacement.
Key results
- MOCID reaches 95.93 AP50 and 98.22 F1 on DAUB, and 94.74 AP50 and 97.88 F1 on IRDST, outperforming the listed baselines.
- Ablation on DAUB: base 83.59 AP50; adding FISTA gives 92.42 and adding DAM gives 90.22; both together give 95.93.
- Input clip length T=5 gives the best AP50 (95.93) on DAUB among T=3 to 9, with inference time rising with T.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FISTA and DAM to the spatial-only CSPDarknet baseline (ablation) | MOCID · DAUB | AP50: 83.59 -> 95.93 | Helped |
| Replace FISTA blocks with convolution blocks (ablation) | CSPDarknet baseline on DAUB · DAUB | AP50: 83.59 -> 85.43 | Helped |
| Add FISTA (ablation) | CSPDarknet baseline on DAUB · DAUB | AP50: 83.59 -> 92.42 | Helped |
| Add DAM (ablation) | CSPDarknet baseline on DAUB · DAUB | AP50: 83.59 -> 90.22 | Helped |
| Use FISTA instead of Tada or TadaCNN (ablation) | MOCID · DAUB | AP50: 91.9; 90.6 -> 92.42 | Helped |
| Add SDS (ablation) | DAM · DAUB | AP50: 93.7 -> 95.9 | Helped |
| Add SDS (ablation) | Video Mamba · DAUB | AP50: 92.5 -> 93.8 | Helped |
| Add SDS (ablation) | Vivim · DAUB | AP50: 93.2 -> 94.7 | Helped |
| Replace DAM with STAM (ablation) | FISTA-based model · DAUB | AP50: 95.93 -> 93.6 | Hurt |
| Set input clip length to 3 rather than 5 (ablation) | MOCID · DAUB | AP50: 95.93 -> 87.57 | Hurt |
| Set input clip length to 6 rather than 5 (ablation) | MOCID · DAUB | AP50: 95.93 -> 95.91 | No clear effect |
| Set input clip length to 7 rather than 5 (ablation) | MOCID · DAUB | AP50: 95.93 -> 94.47 | Hurt |
Surprising findings
- Replacing FISTA blocks with an equal number of plain convolution blocks gives a much smaller gain than FISTA, suggesting the motion context rather than depth drives the improvement.
Limitations
- Inference time grows with clip length T, since T frames are processed simultaneously, making it slower than the single-frame baseline.
- Evaluated only on two MIRSTD datasets (DAUB and IRDST).
Experimental setup
Datasets: DAUB IRDST
Benchmarks: DAUB IRDST
Metrics: AP50 Precision Recall F1
Baselines: DNANet ISNet UIU-Net FC3-Net MTU-Net YOLOx TransVOD++ YOLOv SSTNet LSTFE-Net DTUM Tada
Builds on: Mamba GFNet DTUM PTSEformer CSPDarknet
Closest papers by meaning
- Motion Prior Knowledge Learning with Homogeneous Language Descriptions for Moving Infrared Small Target Detection (AAAI 2025 · similarity 0.86)
- Rethinking Infrared Small Target Detection: A Foundation Driven Efficient Paradigm (NeurIPS 2026 · similarity 0.86)
- Rethinking Evaluation of Infrared Small Target Detection (NeurIPS 2025 · similarity 0.85)
- Pinwheel-shaped Convolution and Scale-based Dynamic Loss for Infrared Small Target Detection (AAAI 2025 · similarity 0.85)
- Semi-supervised Infrared Small Target Detection with Thermodynamic-Inspired Uneven Perturbation and Confidence Adaptation (AAAI 2025 · similarity 0.85)
- SAIST: Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.