🔄 GyroScope — Image Rotation Prediction
GyroScope is a ResNet-18 trained from scratch to detect whether an image is rotated by 0°, 90°, 180°, or 270° — and correct it automatically.
Is that photo upside down? Let GyroScope figure it out.
🎯 Task
Given any image, GyroScope classifies its orientation into one of 4 classes:
| Label | Meaning | Correction |
|---|---|---|
| 0 | 0° — upright ✅ | None |
| 1 | 90° CCW | Rotate 270° CCW |
| 2 | 180° — upside down | Rotate 180° |
| 3 | 270° CCW (= 90° CW) | Rotate 90° CCW |
Correction formula: correction = (360 − detected_angle) % 360
📊 Benchmarks
Trained on 50,000 images from ImageNet-1k × 4 rotations = 200k training samples.
Validated on 5,000 images × 4 rotations = 20k validation samples.
| Metric | Value |
|---|---|
| Overall Val Accuracy | 79.81%% |
| Per-class: 0° (upright) | 79.8% |
| Per-class: 90° CCW | 80.1% |
| Per-class: 180° | 79.4% |
| Per-class: 270° CCW | 79.8% |
| Training Epochs | 12 |
| Training Time | ~4h (Kaggle T4 GPU) |
Benchmark with benchmark.py
===============
RESULTS
===============
Overall result: 411/500 correct
Hit rate: 82.20 %
------------------------------
Details per rotation class:
0° : 96/124 correct ( 77.42%)
90° : 103/119 correct ( 86.55%)
180° : 112/129 correct ( 86.82%)
270° : 100/128 correct ( 78.12%)
==============================
You can also do this benchmark, by using benchmark.py in this repo. :D
Training Curve
| Epoch | Train Acc | Val Acc |
|---|---|---|
| 1 | 41.4% | 43.2% |
| 2 | 52.0% | 46.9% |
| 3 | 59.4% | 62.8% |
| 4 | 64.1% | 66.0% |
| 5 | 67.8% | 69.48% |
| 6 | 70.6% | 72.22% |
| 7 | 73.3% | 74.25% |
| 8 | 75.6% | 76.49% |
| 9 | 77.5% | 77.47% |
| 10 | 79.1% | 79.47% |
| 11 | 80.3% | 79.78% |
| 12 | 80.9% | 79.81% |
🏗️ Architecture
| Detail | Value |
|---|---|
| Base | ResNet-18 (from scratch, no pretrained weights) |
| Parameters | 11.2M |
| Input | 224 × 224 RGB |
| Output | 4 classes (0°, 90°, 180°, 270°) |
| Framework | 🤗 Hugging Face Transformers (ResNetForImageClassification) |
Training Details
- Optimizer: AdamW (lr=1e-3, weight_decay=0.05)
- Scheduler: Cosine annealing with 1-epoch linear warmup
- Loss: CrossEntropy with label smoothing (0.1)
- Augmentations: RandomCrop, ColorJitter, RandomGrayscale, RandomErasing
- ⚠️ No flips — horizontal/vertical flips would corrupt rotation labels
- Mixed precision: FP16 via
torch.cuda.amp
🚀 Quick Start
Installation
pip install transformers torch torchvision pillow requests
Inference — Single Image from URL
python3 use_with_UI.py
--> Download use_with_UI.py first 😄
💡 Example
Input (rotated 180°):
GyroScope Output: 📐 Recognized: 90° | Correction: 270° 📊 Probs: {'0°': '0.0257', '90°': '0.8706', '180°': '0.0735', '270°': '0.0300'} Corrected:
Original Image Source: Link to Pexels
⚠️ Limitations
- Rotationally symmetric images (balls, textures, patterns) are inherently ambiguous — no model can reliably classify these.
- Trained on natural images (ImageNet). Performance may degrade on:
- Documents / text-heavy images
- Medical imaging
- Satellite / aerial imagery
- Abstract art
Only handles 90° increments — arbitrary angles (e.g. 45° or 135°) are not supported! Trained from scratch on 50k images — a pretrained backbone would likely yield higher accuracy (Finetuning).
📝 Use Cases
- 📸 Photo management — auto-correct phone/camera orientation
- 🗂️ Data preprocessing — fix rotated images in scraped datasets
- 🤖 ML pipelines — orientation normalization before feeding to downstream models
- 🖼️ Digital archives — batch-correct scanned/uploaded images
Yesterday, I was sorting photos and like every photo was rotated wrong! This inspired me to make this tool 😂
💻 Training code
The full training code can be found in train.py. Have fun 😊
📜 License
Apache 2.0
🙏 Acknowledgments
- Dataset: ILSVRC/ImageNet-1k
- Architecture: Microsoft ResNet via 🤗 Transformers
- Trained on Kaggle (Tesla T4 GPU)
GyroScope — because every image deserves to stand upright.