File size: 1,997 Bytes
b484048
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
---
license: mit
pipeline_tag: image-segmentation
tags:
- pytorch
- u-net
- swimming-pool
- semantic-segmentation
---

# pool-segmentation: a swimming pool's water

A small U-Net (resnet18 encoder) that finds a swimming pool's **water** in video frames: a class for every
pixel, 0 other, 1 water (swimmers in the water, splashes and lane ropes count as water). About 25 ms per frame on a
MacBook. Made for swimming video analysis (dropping people outside the pool; later, finding the lanes).

Code, training and labeling tools: [github.com/Neptune-Trojans/pool-segmentation](https://github.com/Neptune-Trojans/pool-segmentation) (`pip install pool-segmentation`).

## Use

```python
from pool_segmentation import PoolModel, find_model

model = PoolModel.load(find_model())   # your newest training run in data/models/pool/, else this model
classes = model.predict(frame)         # BGR frame -> class per pixel: 0 other, 1 water
water = model.water(frame)             # the water's outline: one polygon [(x, y), ...] per piece of water
```

The package downloads this model the first time it is needed and keeps it in `~/.cache/huggingface`; when online, it
gets a newer one when there is one. A given version: `find_model("hub:20261119-231345")` (every version is tagged
with its training run).

## This version

| | |
|---|---|
| Training run | `20261119-231345` |
| Classes | 0 other, 1 water |
| Input | frames scaled to 1024×576 (the long side; shape kept) |
| Training | 89 hand-checked frames from 9 videos, 40 epochs |
| Validation | 19 frames (every 6th of each video): water IoU mean 0.996, worst 0.990 |

## Limits

Trained on frames filmed at a few pools, from the deck. Pools that look very different (another camera angle, light,
water color) may need some frames of their own: the package's labeling tools pick frames from your videos, let this
model label them, and you check them in the browser before training again. The training frames are not published:
they show real people.