[Summary] Object detection model "End-to-End Object Detection with Transformers" using Transformer

Introduction

I was curious about "End-to-End Object Detection with Transformers" (DETR), so I read the paper and checked the operation a little. Keep a brief record. [Papers, Github ]

What is DETR (summary)

・ Model released by Facebook AI Research in May this year

・ First use of Transformer, which is famous in the field of natural language processing, for object detection

-Simple network configuration of CNN + Transformer as shown in the figure below

-Achieves "End-to-End" object detection by eliminating parts that require manual adjustment, such as the default values of NMS and AnchorBox.

・ Insist on the effects of "Bipartite Matching Loss" and "Parallel Decoding" as points to realize the above.

・ Applicable not only to object detection but also to segmentation tasks

Inference code example

I will quote the code from the paper. As shown below, the process from model definition to inference processing can be written simply in about 40 lines.

import torch
from torch import nn
from torchvision.models import resnet50

class DETR(nn.Module):

    def __init__(self, num_classes, hidden_dim, nheads,
                 num_encoder_layers, num_decoder_layers):
        super().__init__()
        # We take only convolutional layers from ResNet-50 model
        self.backbone = nn.Sequential(*list(resnet50(pretrained=True).children())[:-2])
        self.conv = nn.Conv2d(2048, hidden_dim, 1)
        self.transformer = nn.Transformer(hidden_dim, nheads,
                                          num_encoder_layers, num_decoder_layers)
        self.linear_class = nn.Linear(hidden_dim, num_classes + 1)
        self.linear_bbox = nn.Linear(hidden_dim, 4)
        self.query_pos = nn.Parameter(torch.rand(100, hidden_dim))
        self.row_embed = nn.Parameter(torch.rand(50, hidden_dim // 2))
        self.col_embed = nn.Parameter(torch.rand(50, hidden_dim // 2))

    def forward(self, inputs):
        x = self.backbone(inputs)
        h = self.conv(x)
        H, W = h.shape[-2:]
        pos = torch.cat([
            self.col_embed[:W].unsqueeze(0).repeat(H, 1, 1),
            self.row_embed[:H].unsqueeze(1).repeat(1, W, 1),
        ], dim=-1).flatten(0, 1).unsqueeze(1)
        h = self.transformer(pos + h.flatten(2).permute(2, 0, 1),
                             self.query_pos.unsqueeze(1))
        return self.linear_class(h), self.linear_bbox(h).sigmoid()

detr = DETR(num_classes=91, hidden_dim=256, nheads=8,
            num_encoder_layers=6, num_decoder_layers=6)
detr.eval()
inputs = torch.randn(1, 3, 800, 1200)
logits, bboxes = detr(inputs)

• If you keep the code of this paper, you cannot load the published trained model. (Because the model definition is written differently) If you want to actually detect with a trained model, it is better to do it like detr_demo.ipynb of the original Github.

Operation check

I actually checked the operation using the trained model. The environment is as follows. ・ OS: Ubuntu 18.04.4 LTS ・ GPU: GeForce RTX 2060 SUPER (8GB) x1 ・ PyTorch 1.5.1 / torchvision 0.6.0

For the implementation of the detection process, we referred to the original detr_demo.ipynb and detected the image captured from the webcam with OpenCV. The model used is a ResNet-50 based DETR.

The following is the actual detection result. It was confirmed that the detection was executed normally. (Although things that are not included in the COCO class are also targeted for shooting)

In my environment, the inference process itself seemed to run in about 45 msec (about 22 FPS).

• Reference values are 28 FPS on V100

end

I studied DETR and confirmed its operation. Since it is a new type of method, I think it will develop further in terms of accuracy and speed. Transformer had the image of being dedicated to natural language processing, but recently it has been introduced into models that handle images. I would like to study more in the future. (I also want to run Image GPT.)

reference

・ "DETR" Transformer object detection debut https://medium.com/lsc-psd/detr-transformer%E3%81%AE%E7%89%A9%E4%BD%93%E6%A4%9C%E5%87%BA%E3%83%87%E3%83%93%E3%83%A5%E3%83%BC-dc18e582dec1 ・ Explanation of End-to-End Object Detection with Transformers (DETR) https://qiita.com/sasgawy/items/61fb64d848df9f6b53d1 -Adopt Transformer for object detection! Detailed explanation of the topic DETR! https://deepsquare.jp/2020/07/detr/