Spaces:
Running
Running
Update README.md
Browse files
README.md
CHANGED
|
@@ -7,4 +7,36 @@ sdk: static
|
|
| 7 |
pinned: false
|
| 8 |
---
|
| 9 |
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
pinned: false
|
| 8 |
---
|
| 9 |
|
| 10 |
+
# LiveCompose
|
| 11 |
+
|
| 12 |
+
LiveCompose is a team project focused on intelligent image cropping for mobile deployment.
|
| 13 |
+
|
| 14 |
+
## Projects
|
| 15 |
+
|
| 16 |
+
### Spaces
|
| 17 |
+
|
| 18 |
+
- **Adacrop Demo**
|
| 19 |
+
Interactive demo for Adacrop, an automatic image cropping system. The demo shows how a trained cropping policy predicts an initial crop box and refines it through sequential actions.
|
| 20 |
+
|
| 21 |
+
## Models
|
| 22 |
+
|
| 23 |
+
- **Adacrop**
|
| 24 |
+
Original PyTorch checkpoints for the AdaCrop model, including the ResNet-50-based Actor-Critic policy and BBox head.
|
| 25 |
+
|
| 26 |
+
- **Adacrop-MNV3-Distilled**
|
| 27 |
+
A distilled MobileNetV3 student version of AdaCrop, designed to reduce model size and make mobile deployment more practical.
|
| 28 |
+
|
| 29 |
+
- **Adacrop-CoreML.mlpackage**
|
| 30 |
+
Core ML exports of Adacrop components for iOS deployment, including BBox prediction and policy action models.
|
| 31 |
+
|
| 32 |
+
## Datasets
|
| 33 |
+
|
| 34 |
+
- **LiveCompose-outpainted-17K**
|
| 35 |
+
Outpainted image dataset with original crop bounding boxes, used for crop prediction and policy distillation experiments.
|
| 36 |
+
|
| 37 |
+
## About AdaCrop
|
| 38 |
+
|
| 39 |
+
AdaCrop formulates image cropping as a two-stage process:
|
| 40 |
+
|
| 41 |
+
1. A BBox head predicts an initial crop region.
|
| 42 |
+
2. An Actor policy refines the crop through actions such as moving, zooming, and stopping.
|