flownet
**FlowNet** is the **pioneering end-to-end deep optical flow architecture that predicts dense motion directly from image pairs** - it demonstrated that learned correspondence can replace many handcrafted motion-estimation stages.
**What Is FlowNet?**
- **Definition**: Convolutional encoder-decoder model for optical flow estimation from two frames.
- **Original Variants**: FlowNetS (stacked input) and FlowNetC (with correlation layer).
- **Output**: Dense per-pixel flow map at full or near-full resolution.
- **Historical Impact**: First major deep model family to make neural flow practical.
**Why FlowNet Matters**
- **End-to-End Learning**: Removed dependence on manual feature engineering in flow pipelines.
- **Speed Gain**: Enabled near real-time flow inference on GPUs at the time.
- **Benchmark Shift**: Sparked rapid progress in learned correspondence models.
- **Architecture Foundation**: Influenced PWC-Net, RAFT, and later refinement approaches.
- **Practical Utility**: Widely reused as initialization and baseline in motion tasks.
**FlowNet Architecture Highlights**
**Encoder-Decoder Core**:
- Downsample to capture large-context matching cues.
- Upsample with skip connections for fine motion recovery.
**Correlation Module (FlowNetC)**:
- Explicitly compares feature patches between two frames.
- Improves matching quality for structured motion.
**Multi-Scale Supervision**:
- Predict flow at several scales and refine progressively.
- Stabilizes training and detail recovery.
**How It Works**
**Step 1**:
- Feed frame pair through convolutional encoder to build feature hierarchy and correspondence cues.
**Step 2**:
- Decode coarse-to-fine flow estimates with skip fusion to produce dense final output.
FlowNet is **the foundational deep optical-flow milestone that proved neural networks can learn pixel correspondence directly from data** - it remains an essential reference point in modern motion estimation history.