Human-Like Delicate Region Erasing Strategy for Weakly Supervised Detection
Authors: Qing En, Lijuan Duan, Zhaoxiang Zhang, Xiang Bai, Yundong Zhang3502-3509
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Consequently, the proposed method is validated on the PASCAL VOC 2007 and PASCAL VOC 2012 datasets. The experimental results show that the proposed method is capable of locating a single object within 5 steps and has great significance to the research on weakly supervised localization with a human-like mechanism. We evaluate the proposed method on the PASCAL VOC 2007 and 2012 datasets. Average precision (AP) and correct localization Cor Loc are used to evaluate the performance of our method. |
| Researcher Affiliation | Collaboration | Qing En,1 Lijuan Duan,1 Zhaoxiang Zhang,2 Xiang Bai,3 Yundong Zhang4 1Beijing Key Laboratory of Trusted Computing, Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China 2Center for Research on Intelligent Perception and Computing, National Laboratory of Pattern Reconition, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China 3School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan 430074, China 4State Key Lab of Digital Multimedia Chip Technology, Vimicro Corp, Beijing 100191, China |
| Pseudocode | Yes | Algorithm 1 Training process of the proposed method Require: Training data D={(xi, yi)}N i=1, replay counter, class num, max epoch, T; 1: Train the feature extractor f with D; 2: for c = 1, class num do 3: Initialize agent DQNc with param θc; 4: for epoch=1,max epoch do 5: for xi,yi in D and c in yi do 6: Set status=1; 7: Get state sxi t , current step confidence clsxi t ; 8: while t < T and status==1 do 9: Calculate M c i,t(u, v) by CAM(xi t, CLS, yi); 10: Select action axi t with ϵ-greedy; 11: if a==6 do 12: status=0; 13: Obtain Maskc xi,t and erasexi t ; 14: Do erase xc i,t+1 = xc i,t Maskc xi,t; 15: Get sxi t+1 and clsxi t+1; 16: Calculate reward rxi t by Eq. 4,5; 17: Store transition(sxi t , axi t , rxi t , sxi t+1); 18: Update confidence clsxi t = clsxi t+1; 19: Do memory replay; 20: end for Ensure: Trained DQN DQN = {DQNc}class num c=1 |
| Open Source Code | No | The paper does not provide concrete access to source code (specific repository link, explicit code release statement, or code in supplementary materials) for the methodology described in this paper. |
| Open Datasets | Yes | We evaluate the proposed method on the PASCAL VOC 2007 and 2012 datasets. |
| Dataset Splits | Yes | We evaluate the proposed method on the PASCAL VOC 2007 and 2012 datasets. PASCAL VOC 2012 val set (Table 2 caption). PASCAL VOC 2007 trainval set (Table 3 caption). |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper mentions software components and frameworks like 'Resnet-50', 'deep Q-network', and 'stochastic gradient descent (SGD)' but does not provide specific version numbers for these or other ancillary software dependencies. |
| Experiment Setup | Yes | We train the network with stochastic gradient descent (SGD) with momentum at a learning rate of 0.01 for 20 epochs. Second, the image descriptor and the history vector are the inputs of the deep Q-network, whose structure is composed of two fully connected layers with 1024 neurons and an action layer with 6 neurons. The architecture of the deep Q-network is shown in Fig. 3. Furthermore, we train each deep Q-network model for 50 epochs, apply an experience replay of 1000 memory capacity, and set the target parameter update step to 100. ... We set γ to 0.9 in our experiments. ... σ is the classification reward, which we set to 3.2 in our experiments. We constrain the classification confidence to not less than ξ = 0.4. ... τ as -0.1 in Eq. 4. ... where we set µ to 0.5 and ψ to 0.8. The erasing degree reward β is set to 0.5 in our experiments. ... where ζ is -0.5 in our experiments. ... The total number of steps T is set to 5. |