Research map / Video / Video segmentation and tracking
Compressive Quantization for Fast Object Instance Search in Videos
Tan Yu, Zhenzhen Wang, Junsong Yuan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| jointly optimize quantization and hashing instead of k-means followed by ITQ | compressive quantization · Groundhog Day (k=40, l=256) | mAP: 0.523 -> 0.654 | Improved |
| jointly optimize quantization and hashing instead of k-means followed by ITQ | compressive quantization · CNN2h (k=40, l=256) | mAP: 0.578 -> 0.751 | Improved |
| binarize k-means codewords with ITQ (Baseline3) instead of keeping real-valued codewords | k-means codewords · CNN2h (k=100) | mAP: 0.845 -> 0.687 | Hurt |
| use compressive quantization instead of real-valued k-means codewords (Baseline2) | compressive quantization · CNN2h (k=100, l=512) | mAP: 0.846 -> 0.847 | No clear effect |
| compress proposals into k=40 k-means codewords instead of exhaustive matching | Baseline1 · CNN2h | mAP: 0.856 -> 0.817 | Mixed |
| increase codewords per video from 10 to 100 (ablation) | Baseline2 · CNN2h | mAP: 0.625 -> 0.845 | Improved |
| use binary codewords of 512 bits instead of real-valued 256-d codewords | Baseline2 · CNN2h (k=100) | memory: 70.32 MB -> 4.39 MB | Improved |
| use Hamming distance on binary codes instead of real-valued distance | Baseline2 · CNN2h (k=100) | search time: 314 milliseconds -> 8 milliseconds | Improved |
| use compressive quantization instead of exhaustive search | Baseline1 · CNN2h | search time and memory: 10.48 s -> 0.008 s; 2.060 GB -> 4.39 MB | Improved |
| use compressive quantization instead of k-means + CBE hashing | k-means + CBE · CNN2h (k=100, l=512) | mAP: 0.791 -> 0.847 | Improved |
| use compressive quantization instead of k-means + LSH | k-means + LSH · Groundhog Day (k=100, l=128) | mAP: 0.393 -> 0.635 | Improved |
| longer binary code from 128 to 512 bits (ablation) | compressive quantization · CNN2h (k=100) | mAP: 0.347 -> 0.847 | Improved |
Closest papers by meaning
- Approximate Search with Quantized Sparse Representations (ECCV 2016 · similarity 0.88)
- HOPE: Hierarchical Object Prototype Encoding for Efficient Object Instance Search in Videos (CVPR 2017 · similarity 0.88)
- Unsupervised Neural Quantization for Compressed-Domain Similarity Search (ICCV 2019 · similarity 0.87)
- Scalable Feature Matching by Dual Cascaded Scalar Quantization for Image Retrieval (TPAMI 2016 · similarity 0.86)
- Central Similarity Quantization for Efficient Image and Video Retrieval (CVPR 2020 · similarity 0.85)
- Deep Visual-Semantic Quantization for Efficient Image Retrieval (CVPR 2017 · similarity 0.85)
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.
