Research map / Detection and segmentation / Tracking
WebUOT-1M: Advancing Deep Underwater Object Tracking with A Million-Scale Benchmark
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.
Introduces WebUOT-1M, a million-frame underwater object tracking benchmark with language prompts, and OKTrack, an omni-knowledge distillation tracker that transfers open-air knowledge from a teacher Transformer to an underwater student ViT.
Problem
Underwater object tracking lacks large, diverse public datasets, and open-air trackers degrade on underwater videos due to domain gaps and sample imbalance with open-air data.
Approach
Build a million-scale UOT benchmark with 408 categories, language prompts and attributes, and distill open-air knowledge from a multi-view teacher (underwater plus enhanced frames) into a plain ViT student using token contrastive, similarity, feature and response distillation losses, plus a training-free Kalman-based motion-aware prediction module.
Key results
- OKTrack retrained on WebUOT-1M reaches 57.5 Pre / 60.0 AUC on WebUOT-1M, compared with 52.9/56.5 for original OSTrack.
- Full OKTrack ablation on WebUOT-1M improves AUC from 56.5 (base) to 58.9 with all distillation losses and MATP (base 52.0/56.5 Pre/AUC).
- Thirty deep trackers are benchmarked cross-domain; Transformer-based and UOT-specific trackers (OKTrack, UOSTrack) rank highest.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CKD only (ablation) | OKTrack base · UTB180 | Pre: 62.3 -> 63.6 | Helped |
| add CKD only (ablation) | OKTrack base · WebUOT-1M | Pre: 52.0 -> 53.9 | Helped |
| add SKD only (ablation) | OKTrack base · UTB180 | Pre: 62.3 -> 63.2 | Helped |
| add SKD only (ablation) | OKTrack base · WebUOT-1M | Pre: 52.0 -> 53.5 | Helped |
| add FKD only (ablation) | OKTrack base · UTB180 | Pre: 62.3 -> 63.2 | Helped |
| add FKD only (ablation) | OKTrack base · WebUOT-1M | Pre: 52.0 -> 53.2 | Helped |
| add RKD only (ablation) | OKTrack base · UTB180 | Pre: 62.3 -> 65.0 | Helped |
| add RKD only (ablation) | OKTrack base · WebUOT-1M | Pre: 52.0 -> 54.8 | Helped |
| add the full CKD, SKD, FKD, and RKD distillation suite (ablation) | OKTrack base · UTB180 | Pre: 62.3 -> 65.4 | Helped |
| add the full CKD, SKD, FKD, and RKD distillation suite (ablation) | OKTrack base · WebUOT-1M | Pre: 52.0 -> 55.2 | Helped |
| add MATP at inference (ablation) | OKTrack with four distillation losses · UTB180 | Pre: 65.4 -> 66.0 | Helped |
| add MATP at inference (ablation) | OKTrack with four distillation losses · WebUOT-1M | Pre: 55.2 -> 56.1 | Helped |
Surprising findings
- Vision-only OKTrack outperforms vision-language trackers such as CiteTracker-256, All-in-One and UVLTrack on WebUOT-1M.
- Language-prompt-only trackers achieve poor results (AUC about 31) on WebUOT-1M.
Limitations
- Relies on ViT backbone with quadratic self-attention cost.
- Dataset size still small compared with large multi-modal datasets such as LAION-5B and InternVid.
- Language annotations describe mostly the initial state and may be ambiguous in long videos.
Experimental setup
Datasets: WebUOT-1M LaSOT GOT-10k TrackingNet COCO OTB99-L TNL2K WebUAV-3M VisualGenome RUOD FishExtend UTB180
Benchmarks: WebUOT-1M UTB180 UOT100 VMAT UVOT400
Metrics: Pre nPre AUC cAUC mACC
Baselines: SiamFC ATOM OSTrack CiteTracker-256 All-in-One UVLTrack UOSTrack SeqTrack-B256
Builds on: knowledge distillation UOSTrack HDETrack OSTrack SORT
Closest papers by meaning
- Diving into Underwater: Segment Anything Model Guided Underwater Salient Instance Segmentation and A Large-scale Dataset (ICML 2024 · similarity 0.85)
- UVEB: A Large-scale Benchmark and Baseline Towards Real-World Underwater Video Enhancement (CVPR 2024 · similarity 0.85)
- WaterDiffusion: Learning a Prior-involved Unrolling Diffusion for Joint Underwater Saliency Detection and Visual Restoration (AAAI 2025 · similarity 0.85)
- NAUTILUS: A Large Multimodal Model for Underwater Scene Understanding (NeurIPS 2025 · similarity 0.85)
- Atlantis: Enabling Underwater Depth Estimation with Stable Diffusion (CVPR 2024 · similarity 0.84)
- VastTrack: Vast Category Visual Object Tracking (NeurIPS 2024 · 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.