Merge pull request #56 from Giskard-AI/readme-image-classification-ob…

…ject-detection

Add readme for image classification and object detection

.github/workflows/do-release.yml

@@ -13,11 +13,16 @@ jobs:
     steps:
       - name: Checkout code
         uses: actions/[email protected]
+
+      - name: Remove dark theme logo from README
+        run:  sed -i 's/.*#gh-dark-mode-only.*//' README.md
+
       - name: Setup PDM
         uses: pdm-project/setup-pdm@v3
         with:
           python-version: "3.10"
           cache: false
+
       - name: Build dist
         run: pdm build
 

README.md

@@ -27,7 +27,8 @@
 ## Supported computer vision tasks
 
 - **Facial Landmark Detection** ([Readme](https://github.com/Giskard-AI/giskard-vision/blob/main/giskard_vision/landmark_detection/Readme.md))
-- **Image Classification** 🔜 Coming soon! 
+- **Image Classification** ([Readme](https://github.com/Giskard-AI/giskard-vision/blob/main/giskard_vision/image_classification/Readme.md))
+- **Object Detection** ([Readme](https://github.com/Giskard-AI/giskard-vision/blob/main/giskard_vision/object_detection/Readme.md))
 
 ## Install
 

giskard_vision/image_classification/Readme.md

@@ -0,0 +1,34 @@
+# `image_classification` module
+
+This module contains model wrappers, dataloaders, tests and all the ingredients needed to evaluate your single-label image classification models.
+In particular this module allows you to evaluate your model against the following criteria:
+
+- Performance on images with different basic image attributes.
+- Performance on images with various metadata from the datasets.
+- Robustness against image perturbations like blurring, resizing, recoloring (performed by `opencv`: https://github.com/opencv/opencv)
+
+## Wrapped Datasets
+
+- [geirhos_conflict_stimuli](https://www.tensorflow.org/datasets/catalog/geirhos_conflict_stimuli) through Tensorflow Datasets
+- [CIFAR100](https://huggingface.co/datasets/uoft-cs/cifar100) through Hugging Face
+- [Skin cancer](https://huggingface.co/datasets/marmal88/skin_cancer) through Hugging Face
+
+
+## Scan and Supported Classification
+
+Once the model and dataloader (`dl`) are wrapped, you can scan the model with the scan API in Giskard vision core:
+
+```python
+from giskard_vision.core.scanner import scan
+
+results = scan(model, dl)
+```
+
+It adapts the [scan API in Giskard Python library](https://github.com/Giskard-AI/giskard#2--scan-your-model-for-issues) to magically scan the vision model with the dataloader.
+
+Currently, due to the constraint of the scan API, we support a subset of image classification tasks:
+
+- [x] Multiclass and single label
+- [ ] Multiclass and multi-label
+
+We will be working to remove such limit for the scan.

giskard_vision/landmark_detection/Readme.md

@@ -22,7 +22,17 @@ You can also check our publicly hosted versions of these datasets on S3:
   - https://poc-face-aligment.s3.eu-north-1.amazonaws.com/300W/300w.zip.004
 - FFHQ (only meta data): https://poc-face-aligment.s3.eu-north-1.amazonaws.com/ffhq/json.zip
 
-## Metrics
+## Scan and Metrics
+
+Once the model and dataloader (`dl`) are wrapped, you can scan the model with the scan API in Giskard vision core:
+
+```python
+from giskard_vision.core.scanner import scan
+
+results = scan(model, dl)
+```
+
+It adapts the [scan API in Giskard Python library](https://github.com/Giskard-AI/giskard#2--scan-your-model-for-issues) to magically scan the vision model with the dataloader. The considered metrics are:
 
 - [x] ME: Mean Euclidean distances
 - [x] NME: Normalised Mean Euclidean distances

giskard_vision/object_detection/Readme.md

@@ -0,0 +1,31 @@
+# `object_detection` module
+
+This module contains model wrappers, dataloaders, tests and all the ingredients needed to evaluate your object detection models.
+In particular this module allows you to evaluate your model against the following criteria:
+
+- Performance on images with different basic image attributes.
+- Performance on images with various metadata from the datasets.
+- Robustness against image perturbations like blurring, resizing, recoloring (performed by `opencv`: https://github.com/opencv/opencv)
+
+## Wrapped Datasets
+
+- [Racoon](https://www.kaggle.com/datasets/debasisdotcom/racoon-detection/data)
+- [300W](https://ibug.doc.ic.ac.uk/resources/300-W/), using the boundary box around all face landmarks
+- [ffhq](https://github.com/DCGM/ffhq-features-dataset), using the boundary box around all face landmarks
+- [Living room passes](https://huggingface.co/datasets/Nfiniteai/living-room-passes) through Hugging Face
+
+## Scan and Metrics
+
+Once the model and dataloader (`dl`) are wrapped, you can scan the model with the scan API in Giskard vision core:
+
+```python
+from giskard_vision.core.scanner import scan
+
+results = scan(model, dl)
+```
+
+It adapts the [scan API in Giskard Python library](https://github.com/Giskard-AI/giskard#2--scan-your-model-for-issues) to magically scan the vision model with the dataloader. The considered metric is:
+
+- [x] Intersection over Union (IoU)
+
+Currently, we only support one object both in the model prediction and ground truth, due to the constraint of the scan API. We will be working to remove such limit.