Deformable DETR: Deformable Transformers for End-to-end Object Detection

URL

• paper: https://arxiv.org/pdf/2010.04159.pdf
• code: https://github.com/fundamentalvision/Deformable-DETR

TL;DR

• 提出了 Deformable DETR：这是一种新的目标检测模型，解决了现有 DETR 模型的收敛速度慢和特征空间分辨率有限的问题。
• 使用可变形的注意力模块：这些模块只关注参考点周围的一小部分关键采样点，从而在更少的训练周期内提高了性能，尤其是对小对象的检测。
• 结合了可变形卷积的稀疏空间采样和 Transformer 的关系建模能力：这使得模型能够在处理大规模数据时保持高效，同时还能捕捉到复杂的上下文关系。
• 引入了一种两阶段的变体：在这个变体中，区域提议由 Deformable DETR 生成，然后进行迭代的细化。这使得模型能够更精确地定位和识别目标。

Algorithm

Deformable DETR 整体结构图

Deformabel Attention Block

• Multi-Head Attention:
  $MultiHeadAtten(z_q, x) = \sum_{m=1}^MW_m[\sum_{k\in\Omega_k}A_{mqk}\cdot W_m'x_k]$

  • 输入为一个 query 的表征 $z_q$ ，以及总特征 x，输出为 query 查询结果向量

  • M 表示 number of head

  • $A_{mqk}$ 表示 $softmax(\frac{QK^T}{\sqrt{d}})$

  • $W_m'x_k$ 实际上就是 self-attention 中的 $V$

• Deformable Attention:
  $DeformableAtten(z_q,p_q,x) = \sum_{m=1}^MW_m[\sum_{k=1}^KA_{mqk}\cdot W_m'x(p_q + \Delta p_{mqk})]$
  • 输入为一个 query 的表征 $z_q$ ，总特征 x，以及 query 对应的 预设采样位置，输出为 query 查询结果向量
  • $\Delta p_{mqk}$ 表示由 $z_q$ 计算得到的 基于预设查询位置的横纵偏移
  • $A_{mqk} = softmax(z_qW_a)\ \ ,W_a\in\mathbb{R}^{dim\times num\_points}\ \ ,z_q\in\mathbb{R}^{dim}$ ，即 point position attention 是由 query 线性映射得到的 ，因此 Deformable Attention 没有 Key 的存在，只有 Query 和 Value
  • K 表示 number of points，即采样点个数
• Multi-Scale Deformable Attention：
  $MSDeformableAtten(z_q,\hat{p}_q,\{x\}_{l=1}^L) = \sum_{m=1}^MW_m[\sum_{l=1}^L\sum_{k=1}^KA_{mlqk}\cdot W_m'x^l(\phi_l(\hat{p}_q) + \Delta p_{mlqk})]$
  • 与 Deformable Attention 不同的是，输入的 x 变成了多尺度特征（例如 backbone 不同深度的特征），更贴近实际视觉工程化应用场景
  • point 采样范围是所有 level 的 feature map，即 MSDefromableAttention 有全局 attention 信息
• Deformable Attention 和 Self Attention 对比

import torch
import torch.nn.functional as F
import math
def attention(query, key, value):
    # query, key, value shapes: (batch_size, sequence_length, embedding_dim)
    scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(query.size(-1))
    # scores shape: (batch_size, sequence_length, sequence_length)
    probs = F.softmax(scores, dim=-1)
    # probs shape: (batch_size, sequence_length, sequence_length)
    output = torch.matmul(probs, value)
    # output shape: (batch_size, sequence_length, embedding_dim)
    return output
def deformable_attention(query, value, reference_points, num_sampling_points, atten_linear, offset_linear):
    # query shape: (batch_size, sequence_length, embedding_dim)
    # value shape: (batch_size, sequence_length, embedding_dim)
    # reference_points shape: (batch_size, sequence_length, 2)
    # num_sampling_points: integer, number of points to sample around each reference point
    
    batch_size, seq_len, embed_dim = query.size()
    
    # Calculate offsets
    # offset_linear is a linear layer that predicts the offsets
    offsets = offset_linear(reference_points).view(batch_size, seq_len, num_sampling_points, 2)
    # offsets shape: (batch_size, sequence_length, num_sampling_points, 2)
    
    # Calculate sampling positions based on reference points
    sampling_positions = reference_points.unsqueeze(2) + offsets
    # sampling_positions shape: (batch_size, sequence_length, num_sampling_points, 2)
    
    # Sample values (this is simplified; you might need interpolation)
    # Here, we assume value and reference_points are in the same space for simplicity
    sampling_values = value.gather(1, sampling_positions.long())
    # sampling_values shape: (batch_size, sequence_length, num_sampling_points, embedding_dim)
    
    # Calculate scores
    # atten_linear is a linear layer that transforms the query for calculating attention scores
    scores = atten_linear(query).view(batch_size, seq_len, num_sampling_points)
    # scores shape: (batch_size, sequence_length, num_sampling_points)
    
    # Softmax to get attention probabilities
    probs = F.softmax(scores, dim=-1)
    # probs shape: (batch_size, sequence_length, num_sampling_points)
    
    # Calculate output
    output = torch.matmul(probs, sampling_values)
    # output shape: (batch_size, sequence_length, embedding_dim)
    
    return output

Thought

• 用 query 线性映射代替 query 和 key 外积做 attention 数学上可解释性会变差，计算复杂度会降低
• Deformable Conv 是典型的对 NPU 不友好，Deformable Attention 会更复杂，被代季峰支配的恐惧
• 用 Multi-scale 做各特征尺度上的信息融合，开创了一个 CNN 做 backbone + Deformable Transformer 做 head 的计算机视觉任务模型新范式，甚至省去了 FPN
• 总之是用各种便宜的计算来近似复杂的全局 attention，复杂度从 H*W --> K，即 $O(n^2) -> O(K)$