svhn-detection-tf

Object detection on SVHN dataset in tensorflow using efficientdet

Getting Started

First of all you need to call pip install git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI since there is no official PiPy version. Please verify the validity of your changes before pushing by running $ python train.py --test.

Sanity Checks

The classification loss should start at ~6.0 and should get to at most ~0.1. The regression loss should start at ~0.3 and should get to ~0.07. One epoch should take ~1 minute.

Experiments

TODO: add a table with results

Tasks

Tasks should be completed in chronological order.

Verify correct implementation of Straka's metric

Verify that predictions are computed correctly (it returns a sane value) and the val_score is computed correctly from the obtained predictions.

Implement COCO metrics

COCO metric should validate the performance of the model during training and after the training finished. The implementation could be based on https://github.com/google/automl/blob/master/efficientdet/coco_metric.py and could used the predict function (similar to straka metric). In the fit function, the predictions are already computed. The metric could use pycocotools.

Implement augmentations from efficientdet paper

Augmentations should be the following (if I remember correctly): translation, scale. Bounding boxes should be transformed accordingly. The correct behaviour should be verified in jupiter with bounding boxes drawn over the picture

Implement autoaugment

Follow https://github.com/google/automl/blob/master/efficientdet/aug/autoaugment.py

Experiment with different image_size

Make sure, that the smallest conv layer in efficientdet has at least 4 pixels. Increase/decrease number of efficientdet layers if needed

Experiment with different efficientdet architecture

Try using more/less features from the efficientdet or even replace the efficientdet with a single output and map it to multi-sized anchors.

Experiment with different learning rate/wd/momentum/grad_clip

Compare with https://github.com/penny4860/retinanet-digit-detector

Finetune object detection

Experiment with iouthreshold and scorethreshold to obtain maximal performance out of the network. Also finetune/replace the combined non-maximum suppression (read the docs).

Evaluate on Straka's test set

Implement Stochastic Depth

!!! stochastic depth is not implemented in the efficientdet paper!!!

Finetune performance

Computing metrics could be done in better way mAP could be implemented as a keras metric Data pipeline could be optimized