State-of-the-art Computer Vision and Object Detection for TensorFlow.
Made by Rishabh Anand • https://rish-16.github.io
sightseer provides state-of-the-art general-purpose architectures (YOLOv3, MaskRCNN, Fast/Faster RCNN, SSD...) for Computer Vision and Object Detection tasks with 30+ pretrained models written in TensorFlow 1.15.
sightseer is written in Python 3.5+ and TensorFlow 1.15.
sightseer, you must first have TensorFlow installed. To do so, follow the instructions on the TensorFlow installation page.
When your virtual environment is set up with TensorFlow, you can install
pip install sightseer
Model Clients (as of now)
YOLOv3Client(Darknet by Joseph Redmon)
Facebook AI's Mask R-CNN will be out in the v1.2.0 release. For more information on model release, check out the Roadmap.
The package comes with 4 major components that help with different parts of the object detection process all the way from preparing your raw data to getting predictions and displaying them.
|Sightseer||Obtains image data or video footage|
|Proc||Provides image/frame-wise annotation and inter-format conversion tools|
|Zoo||Stores the wrappers over all state-of-the-art models and configs|
|Serve||Provides deployment and model serving protocols and services|
If not using custom datasets,
Zoo are the submodules majorly used for generic predictions from pre-trained models. When there is custom data involved, you can use
Proc to annotate your datasets and even convert them between XML/JSON/CSV/TFRecord formats.
Serveis an experimental productionising submodule that helps deploy your models on cloud services like AWS and GCP. For more details on future tools and services, check out the Roadmap.
Footage or raw images can be rendered using
Sightseer before being ingested into models or further preprocessed.
1a. Loading images
from sightseer import Sightseer ss = Sightseer() image = ss.load_image("path/to/image") # return numpy array representation of image
1b. Loading videos
from sightseer import Sightseer ss = Sightseer() frames = ss.load_vidsource("path/to/video") # returns nested array of frames
Support for video, webcam footage, and screen recording will be out in the coming v1.2.0 release.
2. Using models from
Once installed, any model offered by
sightseer can be accessed in less than 10 lines of code. For instance, the code to use the YOLOv3 (Darknet) model is as follows:
from sightseer import Sightseer from sightseer.zoo import YOLOv3Client yolo = YOLOv3Client() yolo.load_model() # downloads weights # loading image from local system ss = Sightseer() image = ss.load_image("./assets/road.jpg") # getting labels, confidence scores, and bounding box data preds, pred_img = yolo.predict(image, return_img=True) ss.render_image(pred_img)
To run the model on frames from a video, you can use the
from sightseer import Sightseer from sightseer.zoo import YOLOv3Client yolo = YOLOv3Client() yolo.load_model() # downloads weights # loading video from local system ss = Sightseer() frames = ss.load_vidsource("./assets/video.mp4") """ For best results, run on a GPU """ # getting labels, confidence scores, and bounding box data preds, pred_frames = yolo.framewise_predict(frames) ss.render_footage(pred_frames) # plays the video and saves the footage
The module can even be repurposed into a Command-line Interface (CLI) app using the
Suggestions, improvements, and enhancements are always welcome! If you have any issues, please do raise one in the Issues section. If you have an improvement, do file an issue to discuss the suggestion before creating a PR.
All ideas – no matter how outrageous – welcome!
Before committing, please check the Roadmap to see if proposed features are already in-development or not.
Note: Please commit all changes to the
developmentexperimentation branch instead of