Research map / Video / Action recognition
Massively Parallel Video Networks
Joao Carreira, Viorica Patraucean, Laurent Mazare, Andrew Zisserman, Simon Osindero
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 |
|---|---|---|---|
| distill activations from a sequential teacher into the parallel model (ablation) | Par-Inception parallel (10 subnets) · miniKinetics action recognition | top-1 accuracy: 54.5 -> 61.2 | Improved |
| distill activations from a sequential teacher (ablation) | Par-DenseNet with clocks · miniKinetics keypoint localisation | weighted sigmoid cross-entropy | Improved |
| increase parallelism from sequential to 10 subnets (ablation) | Par-Inception · miniKinetics action recognition | top-1 accuracy: 71.4 -> 54.5 | Hurt |
| increase parallelism from sequential to 14 subnets (ablation) | Par-DenseNet · miniKinetics action recognition | top-1 accuracy: 67.6 -> 54.0 | Hurt |
| add multi-rate clocks to sequential model (ablation) | Par-DenseNet · Throughput 8 GPUs | speedup: 1.0 -> 3.4 | Improved |
| add multi-rate clocks to fully parallel model (ablation) | Par-DenseNet · Throughput 8 GPUs | speedup: 3.8 -> 7.4 | Improved |
| add multi-rate clocks to fully parallel model (ablation) | Par-Inception · Throughput 8 GPUs | speedup: 2.6 -> 5.0 | Improved |
| add multi-rate clocks (ablation) | Par-DenseNet · miniKinetics keypoint localisation | weighted sigmoid cross-entropy | Hurt |
| add temporal filters (ablation) | Par-DenseNet · miniKinetics keypoint localisation | weighted sigmoid cross-entropy | Mixed |
| add feedback connection (ablation) | Par-DenseNet · miniKinetics keypoint localisation | weighted sigmoid cross-entropy | Improved |
| train with depth-parallel connectivity instead of rewiring a sequential model (ablation) | Par-DenseNet · miniKinetics keypoint localisation | test loss | Improved |
| allow prediction latency of 2 frames (ablation) | Par-DenseNet up to 4 subnets · miniKinetics keypoint localisation | test loss | Improved |
Closest papers by meaning
- AssembleNet: Searching for Multi-Stream Neural Connectivity in Video Architectures (ICLR 2020 · similarity 0.86)
- ECO: Efficient Convolutional Network for Online Video Understanding (ECCV 2018 · similarity 0.85)
- A Coding Framework and Benchmark Towards Low-Bitrate Video Understanding (TPAMI 2024 · similarity 0.85)
- Generating, Fast and Slow: Scalable Parallel Video Generation with Video Interface Networks (ICCV 2025 · similarity 0.85)
- VideoLucy: Deep Memory Backtracking for Long Video Understanding (NeurIPS 2025 · similarity 0.84)
- CoPE-VideoLM: Leveraging Codec Primitives For Efficient Video Language Modelling (NeurIPS 2026 · 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.
