DETR (End-to-End Object Detection) model with ResNet-50 backbone trained on SKU110K Dataset with 400 num_queries

DEtection TRansformer (DETR) model trained end-to-end on SKU110K object detection (8k annotated images) dataset. Main difference compared to the original model is it having 400 num_queries and it being pretrained on SKU110K dataset.

How to use

Here is how to use this model:

from transformers import DetrImageProcessor, DetrForObjectDetection
import torch
from PIL import Image, ImageOps
import requests

url = "https://github.com/Isalia20/DETR-finetune/blob/main/IMG_3507.jpg?raw=true"
image = Image.open(requests.get(url, stream=True).raw)
image = ImageOps.exif_transpose(image)

# you can specify the revision tag if you don't want the timm dependency
processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50", revision="no_timm")
model = DetrForObjectDetection.from_pretrained("isalia99/detr-resnet-50-sku110k")
model = model.eval()
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)

# convert outputs (bounding boxes and class logits) to COCO API
# let's only keep detections with score > 0.8
target_sizes = torch.tensor([image.size[::-1]])
results = processor.post_process_object_detection(outputs, target_sizes=target_sizes, threshold=0.8)[0]

for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
        box = [round(i, 2) for i in box.tolist()]
        print(
                f"Detected {model.config.id2label[label.item()]} with confidence "
                f"{round(score.item(), 3)} at location {box}"
        )

This should output:

Detected LABEL_1 with confidence 0.983 at location [665.49, 480.05, 708.15, 650.11]
Detected LABEL_1 with confidence 0.938 at location [204.99, 1405.9, 239.9, 1546.5]
...
Detected LABEL_1 with confidence 0.998 at location [772.85, 169.49, 829.67, 372.18]
Detected LABEL_1 with confidence 0.999 at location [828.28, 1475.16, 874.37, 1593.43]

Currently, both the feature extractor and model support PyTorch.

Training data

The DETR model was trained on SKU110K Dataset, a dataset consisting of 8,219/588/2,936 annotated images for training/validation/test respectively.

Training procedure

Training

The model was trained for 140 epochs on 1 RTX 4060 Ti GPU(Finetuning decoder only) with batch size of 8 and 70 epochs(finetuning the whole network) with batch size of 3 and accumulating gradients for 3 steps.

Evaluation results

This model achieves an mAP of 58.9 on SKU110k validation set. Result was calculated with torchmetrics MeanAveragePrecision class.

Training Code

Code is released in this repository Repo Link. However it's not finalized/tested well yet but the main stuff is in the code.