NaveenKumarGottipalli/FinalShopliftingdetector

Shoplifting_Detection — Multi-Model Shoplifting Detection Benchmark

Overview

• Project 1 (3D CNNs): R(2+1)D baseline vs. DMD-A-Net (Frame-Differential Motion Gate + Temporal Concealment Cross-Attention)
• Project 2 (Video Transformers): TimeSformer baseline vs. TCA-ViT (Motion-Salience Patch Gate + Occlusion Gradient Head)
• Project 3 (Skeleton GCNs, YOLOv26n-pose): ST-GCN vs. CTR-GCN vs. VDA-GCN (Velocity-Gated Temporal Conv + Directed Concealment Adjacency)

Cross-Dataset Evaluation Plan (NaveenKumarGottipalli/Shoplifting_DataCollection)

• Training & validation corpus: S3 + Shop DataSet + UCF-Crime
• Test set: DCSASS Dataset (DCSASS Dataset/Labels/Shoplifting.csv). Step 3B audits source-video overlap with the training corpus (leakage_report.json) and builds a leak-free split (*_lf.json)

Repository structure

• dataset_splits/ — verified sample manifests (train.json, val.json, test.json, verified_all.json, quarantined.json), viability_report.csv, dataset charts and input→output sample figures
• Project_1_R2Plus1D_DMD_ANet/ — results/ (JSON metrics), figures/ (accuracy / loss / ROC / confusion matrix / input-vs-output), figures/limitations/ (16 limitation figures), models/ (trained weights)
• Project_2_TimeSformer_TCA_ViT/ — same layout, 13 limitation figures
• Project_3_Skeleton_GCNs/ — same layout, 8 limitation figures, plus the YOLO pose model in models/
• All_Projects_Comparison/ — cross-project summary table (all_models_summary.csv) and combined ROC / F1 / AUC chart
• MANIFEST.csv — every file with its size and SHA-256

Model weights above GitHub's 100 MB limit are hosted on Hugging Face (NaveenKumarGottipalli/Shoplift_detection), Kaggle and Google Drive instead.

Evaluation protocol (v2)

• dataset_splits/leakage_report.json and dcsass_label_audit.csv document the source-video overlap between the training corpus and DCSASS and the label structure of the DCSASS test set; *_lf.json is the leak-free split.
• All_Projects_Comparison/publication_table_*.csv report AUC / AP with video-clustered bootstrap intervals, validation-tuned thresholds, within-video AUC and calibration; multiseed_results/ holds the per-seed result JSONs.
• Stored v1 figures come from a single seed on the original split; read the v2 tables before citing any number.