Bike Detection Model
Model Description
This model is a lightweight bicycle detector designed for embedded computer vision applications.
It takes a low-resolution grayscale image as input and predicts bicycle bounding boxes using an anchor-free detection head.
- Input shape:
1 × 120 × 160 - Input format: QQVGA grayscale image
- Output shape:
5 × 30 × 40 - Output stride:
4 - Task: Bicycle detection
For each cell of the 30 × 40 output grid, the model predicts:
[L, R, B, T, confidence]
where L, R, B, and T represent the distances from the grid-cell center to the corresponding sides of the predicted bounding box.
The final channel represents the bicycle detection confidence.
Intended Use
The model is intended for:
- Embedded bicycle detection
- Low-power computer vision
- Edge AI and TinyML applications
- Low-resolution grayscale cameras
It was designed with deployment on resource-constrained hardware in mind.
Training Data
The model was trained on a private bicycle detection dataset.
The dataset contains grayscale images prepared for the target embedded use case and includes both positive bicycle samples and negative/background samples.
The training dataset is not publicly distributed with this model.
Model Output
The raw output tensor has shape:
[5, 30, 40]
with the channels corresponding to:
0: L
1: R
2: B
3: T
4: confidence
A post-processing step is required to:
- Decode bounding boxes from the grid predictions.
- Apply a confidence threshold.
- Apply Non-Maximum Suppression (NMS).
Limitations
The model was trained on a private dataset representative of its target application.
Performance may therefore degrade on images with significantly different:
- Camera viewpoints
- Lighting conditions
- Image quality
- Bicycle sizes
- Backgrounds
- Acquisition hardware
The model is intended primarily for the domain represented by its training data.