File size: 2,914 Bytes
4498bd2
 
 
5441555
4498bd2
 
 
5441555
1d3626b
4498bd2
 
5441555
4498bd2
1d3626b
4498bd2
8cf0ff1
1d3626b
8cf0ff1
4498bd2
5441555
 
1d3626b
 
 
5441555
1d3626b
5441555
1d3626b
5441555
8cf0ff1
1d3626b
 
 
5441555
 
 
 
 
 
4498bd2
5441555
 
 
1d3626b
5441555
 
 
 
 
1d3626b
5441555
 
 
 
 
 
 
 
 
4498bd2
 
5441555
 
1d3626b
4498bd2
5441555
4498bd2
5441555
 
 
4498bd2
5441555
 
4498bd2
5441555
 
4498bd2
5441555
 
 
4498bd2
 
 
 
5441555
4498bd2
5441555
4498bd2
5441555
 
 
1d3626b
 
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
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
---
license: mit
tags:
  - target-plp
  - power-lines
  - polyline
  - yolino
  - cross-domain
  - zero-shot
---

# Target-PLP

**Target-PLP** is a cross-domain **test-only** dataset for **zero-shot** evaluation of aerial power-line detection models.

It contains 194 aerial frames annotated with **polyline** instance labels in YOLinO format. Models are trained on [TTPLA](https://huggingface.co/datasets/V8heart/yolino-ttpla-benchmark) and evaluated on Target-PLP **without fine-tuning**.

Compared to TTPLA, Target-PLP is **more complex and challenging**: denser wire layouts, more parallel instances per frame, greater scene variation, and a different imaging domain. Annotations are **ordered polylines** (vertex sequences along each wire), not axis-aligned boxes or masks.

## Statistics

| Split | Images | Resolution | Purpose |
|-------|--------|------------|---------|
| test  | **194** | 512×512 | zero-shot evaluation only |

No train or validation split is provided.

## Preprocessing

1. **Source**: CVAT XML exports from aerial video frames
2. **Resize**: each frame is stretch-resized to 512×512 (LANCZOS); polyline vertices scaled with `sx = 512/W`, `sy = 512/H`
3. **No tiling / no dual-crop** — one full frame per image
4. Frames without valid polyline instances (fewer than 2 points) are skipped

## Directory layout

```
images/test/*.png
labels/test/*.npy
```

Image and label files share the same stem, e.g. `ANAR4318_frame_000000.png``ANAR4318_frame_000000.npy`.

## Annotation format (polyline)

Each `labels/test/<stem>.npy` is a Python dict (**format version 3**):

```python
{
    "polylines": [                    # one polyline per power-line instance
        [[x1, y1], [x2, y2], ...],   # instance 1: ordered vertices (pixels)
        [[x1, y1], [x2, y2], ...],   # instance 2
        ...
    ],
    "instance_ids": [1, 2, ...],      # unique ID per wire instance
    "source_file": "<stem>",
    "image_size_wh": [512, 512],
    "format_version": 3
}
```

- Coordinates are in **pixel space** after the 512×512 resize
- Each polyline is an ordered sequence of 2D points along one wire
- Multiple parallel wires are separate instances with distinct `instance_ids`

## Loading example

```python
import numpy as np
from PIL import Image

root = "./Target-PLP"
stem = "ANAR4318_frame_000000"

img = Image.open(f"{root}/images/test/{stem}.png")          # 512×512 RGB
label = np.load(f"{root}/labels/test/{stem}.npy", allow_pickle=True).item()

polylines    = label["polylines"]       # list[list[[x, y]]]
instance_ids = label["instance_ids"]    # list[int]
```

## Download

```bash
hf download V8heart/Target-PLP \
  --repo-type dataset \
  --local-dir ./Target-PLP
```

## Related

- TTPLA training benchmark: [V8heart/yolino-ttpla-benchmark](https://huggingface.co/datasets/V8heart/yolino-ttpla-benchmark)
- Code: [V8heart/CAPSTONE](https://github.com/V8heart/CAPSTONE)