Research map / Video / Video segmentation and tracking
Weakly Supervised Video Individual Counting
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 weakly supervised video individual counting (WVIC), using only inflow/outflow labels instead of trajectories, and a baseline CGNet trained with a group-level matching soft contrastive loss and a memory-based count predictor, plus a new UAV dataset UAVVIC.
Problem
Video individual counting normally requires expensive per-person trajectory labels to learn identity association across frames; the paper asks whether counting unique individuals can be learned from cheaper identity-agnostic inflow/outflow labels.
Approach
Relax individual-level identity association to group-level matching between adjacent frames: a soft contrastive loss (optimal transport via Sinkhorn) pulls shared individuals together and pushes inflow/outflow individuals apart, and a hinge loss constrains cross-group similarities; inference uses a memory of individual templates with Hungarian matching to count inflows.
Key results
- On SenseCrowd, CGNet reaches MAE 8.86 vs 12.3 for DRNet (about 28% lower).
- On CroHD, CGNet reaches MAE 75.0 vs 141.1 for DRNet, without trajectory labels.
- On UAVVIC, average MAE improves by about 30% over DRNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add constraints on S1 and S2 in the group-level matching loss (ablation) | CGNet · SenseCrowd test | MAE: 13.35 -> 8.86 | Helped |
| add the hinge loss in Eq. (5) (ablation) | CGNet · SenseCrowd test | MAE: 9.84 -> 8.86 | Helped |
| add the memory-based individual count predictor (MCP) (ablation) | CGNet · SenseCrowd test | WRAE (%): 16.9 -> 12.6 | Helped |
| set the sampling interval δ to 4 instead of 3 (ablation) | CGNet · SenseCrowd test | MAE: 8.86 -> 8.71 | Helped |
| replace FIDT with YOLOv8 small as the locator (ablation) | CGNet · SenseCrowd test | MAE: 8.86 -> 7.56 | Helped |
| replace FIDT with YOLOv8 nano as the locator (ablation) | CGNet · SenseCrowd test | MAE: 8.86 -> 8.01 | Helped |
| replace ConvNext-S with PrRoIPooling and replace MCP with DRNet inflow reasoning (ablation) | CGNet · SenseCrowd test | MAE: 10.01 -> 12.59 | Hurt |
| train existing VIC methods with CGNet-generated pseudo trajectory labels instead of ground-truth trajectory labels (ablation) | existing VIC methods on SenseCrowd · SenseCrowd test | MAE: 12.3 -> 14.5 | Hurt |
Surprising findings
- Weakly supervised CGNet outperforms the trajectory-supervised DRNet and MOT methods on all three datasets.
- Training existing VIC methods on CGNet-generated pseudo trajectories causes only slight degradation versus ground-truth trajectories.
Limitations
- Relies on a separate crowd locator (FIDT) trained with coordinate or box annotations.
- Performance is weaker on very high-density scenes such as CroHD15 and extremely dense D4 videos.
- Memory size and time-to-live hyperparameters require tuning.
Experimental setup
Datasets: SenseCrowd CroHD UAVVIC VisDrone
Benchmarks: SenseCrowd CroHD UAVVIC
Metrics: MAE MSE WRAE
Baselines: DRNet FairMOT HeadHunter-T SMILE SparseTrack Deep-OC-SORT BoT-SORT LOI PHDTT
Builds on: DRNet FIDT ConvNeXt Sinkhorn algorithm
Closest papers by meaning
- DR.VIC: Decomposition and Reasoning for Video Individual Counting (CVPR 2022 · similarity 0.86)
- Bootstrapping MLLM for Weakly‑Supervised Class‑Agnostic Object Counting (ICLR 2026 · similarity 0.84)
- Learning to Track Instances without Video Annotations (CVPR 2021 · similarity 0.83)
- Learning to Count without Annotations (CVPR 2024 · similarity 0.82)
- Is ImageNet worth 1 video? Learning strong image encoders from 1 long unlabelled video (ICLR 2024 · similarity 0.82)
- Efficient Masked AutoEncoder for Video Object Counting and A Large-Scale Benchmark (ICLR 2025 · similarity 0.82)
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.