---
language:
- en
tags:
- human-action-recognition
- har
- hog
- cv
- kth
license: mit
---

# KTH HOG Features (NPZ)

This repository hosts a precomputed HOG feature dataset derived from the KTH Human Action Recognition (KTH-HAR) videos. It is used by the CNN-for-HAR Streamlit demo and allows fast inference/training without decoding raw videos at runtime.

## What this file contains

`hog_aug_4framegap.npz` (~3GB) includes:

- `features`: shape `(N, T*3780)` float32  
- `bboxes`: shape `(N, T, 4)` float32, normalized `(cx, cy, w, h)` per frame  
- `labels`: shape `(N,)` int64  
- `metadata`: list of dicts with `video_key`, `subject`, `action`, `split`, `group_idx`, `aug_idx`, `aug_name`, `frame_indices`  
- `config`: dict with generation settings (see below)

Each sample is a clip of **T frames**, stored as a flattened HOG vector. HOG is computed on 64×128 crops (OpenCV HOG defaults), which yields **3780 features per frame**.

---

## How the file is generated (pipeline)

This `.npz` is produced by `extract_hog_augmented.py` in the main repo:

1. Load **bbox metadata JSON** (groups of frame indices per clip).
2. For each video:
   - Read only the needed frames.
3. For each clip:
   - **Always include the original version** (no augmentation).
   - For **train clips only**, generate additional augmented versions (`num_aug - 1`).
4. For every variant:
   - Crop person bbox → resize to 64×128  
   - Optional flip + photometric changes  
   - Compute HOG  
   - Save HOG + normalized bbox info  

---

## What `num_aug=4` means

When `num_aug=4`, each **train clip** becomes **4 variants**:

1. `orig` — no augmentation  
2. `flip` — horizontal flip only  
3. `jit0` — random jitter + photometric + possible blur/noise  
4. `jit1` — another random jitter + photometric + possible blur/noise  

Test clips always have **only the original version**.

---

## Augmentations used in `jit*`

The `jit*` variants apply a random combination of:

- **BBox jitter**: translation `(dx, dy)` + scale  
- **Random flip** (50%)  
- **Photometric changes**: contrast (`alpha`), brightness (`beta`)  
- **Gamma**  
- **Noise** (optional)  
- **Blur** (optional)

### `mild` profile (default)
- scale: 0.92–1.08  
- dx: ±5, dy: ±3  
- alpha: 0.85–1.15  
- beta: -15…+15  
- gamma: 0.95–1.05  
- noise: std 2.0 with p=0.5  
- blur: p=0.15 (kernel 3)

### `strong` profile
- scale: 0.88–1.15  
- dx: ±8, dy: ±6  
- alpha: 0.75–1.25  
- beta: -25…+25  
- gamma: 0.85–1.15  
- noise: std 4.0 with p=0.7  
- blur: p=0.25 (kernel 3)

---

## What `frame_gap=4` means

The **frame gap** is defined in the bbox JSON (not in the extractor).  
Each clip groups frames spaced by a fixed gap (e.g., every 4th frame).  
That’s why this file is named `hog_aug_4framegap.npz`.

---

## Direct download
- https://huggingface.co/datasets/Mihai20/kth-hog-npz/resolve/main/hog_aug_4framegap.npz

---

## Example usage (Python)

```python
import numpy as np

data = np.load("hog_aug_4framegap.npz", allow_pickle=True)
features = data["features"]
bboxes = data["bboxes"]
labels = data["labels"]
metadata = data["metadata"].tolist()
config = data["config"].item()  # contains num_aug, profile, etc.
```

## Notes
- Intended for inference + demo usage without raw video decoding.
- The original videos are not required to use this .npz.
- Augmentation settings are saved in config inside the file.