Faster-rcnn is a family of object detection architectures and models,This repository uses pretrained faster-rcnn model on COCO dataset for detect items in grocery dataset.

---

image2products.json: This file contains predictions of the number of items present in per image.

metrics.json: This file contains scores of evaluation metrics (mAP,precision,recall).

Data preparation:

Data preparation completed in 4 steps:

1. convert csv annotation to pandas DataFrame and separate train and test DataFrames.
2. Fix images which is rotated 180 degree clockwise and 90 degree counter clockwise.
3. convert bounding boxes to float (because Faster-rcnn takes floats in bounding box).
4. Add 1 to all categories because i wand 0 to categorized as is_crowd(i fixed it later).

Augmentation:

I used Horizontal flip with 0.3 probability,Random brightness Contrast with 0.5 probability, To Gray with 0.5 probability and vertical flip with 0.4 probability. I used PascalVOC format for bounding box augmentation.

Detection network:

FasterRCNN with Resnet50 backbone. model is available in torchvision module.

optimizer : AdamW.

learning rate scheduler: StepLR

training parameters:

• batch size: 4,
• Learning rate: 0.0001,
• step-size: 10,
• epochs: 15,
• weight-decay: 0.0001,
• gamma: 0.1

Evaluation:

mean average precision calculated on 0.5 iou-threshold.

Quick Start Examples

---

Python>=3.8.0 is required with all requirements.txt:

$ cd product_detection_sudhanshu_singh
$ pip install -r requirements.txt

Training:

Run commands below to start train model,for training model takes preprocessed data thus data preprocessing automatically covered in this step.

$ python train.py --batch-size 4 --epochs 25 --step-size 9 --lr-rate 0.0001 --weight-decay 0.0001 --gamma 0.1

My prediction scores:

"mAP": 0.6902879417612586,
"precision": 0.7323759327193174,
"recall": 0.9036605018549191

Evaluation:

If you have run train.py then you have got weights for evaluate trained model on test data.Run command below to start evaluation:

if you don't want to train the model then download weights from here

$ python evaluation.py --weights "saved_weights.pth" --iou-thres 0.5

Visualization:

If you have weights then you can view predictions on test images:

# if view prediction on test images
$ python visualization.py --weights "saved_weights.pth" --view-predict
#if view train images
$ python visualization.py --view-train

prediction output on test images: