Rethinking Image Restoration for Object Detection
Authors: Shangquan Sun, Wenqi Ren, Tao Wang, Xiaochun Cao
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments in image dehazing and low light enhancement and show the superiority of our method over conventional training and other domain adaptation and multi-task methods. |
| Researcher Affiliation | Collaboration | Shangquan Sun State Key Laboratory of Information Security Institute of Information Engineering Chinese Academy of Sciences Beijing, China 100093 sunshangquan@iie.ac.cn Wenqi Ren School of Cyber Science and Technology Sun Yat-sen University, Shenzhen Campus Shenzhen, China 518107 rwq.renwenqi@gmail.com Tao Wang Huawei Technologies Co., Ltd. Beijing, China 100085 wangtao10@huawei.com Xiaochun Cao School of Cyber Science and Technology Sun Yat-sen University, Shenzhen Campus Shenzhen, China 518107 caoxiaochun@mail.sysu.edu.cn |
| Pseudocode | Yes | Algorithm 1: Framework of Adam-variant adversarial example generation. Input: Input dataset U with samples x with annotation ˆy, Detector D, Number of attack iteration T, Update stepsize λ, Magnitude tolerance of Perturbation δ Output: Attacked adversarial examples ˆx , |
| Open Source Code | Yes | Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] We include the code in supplemental materials. |
| Open Datasets | Yes | Dehazing Dataset: RTTS [20] is a real-world hazy dataset with detection annotation for test purpose. ... We select those PASCAL VOC images with the above-mentioned five classes and simulate haze on them following [23]. ... Low-light Enhancement Dataset: Ex Dark [26] is a natural low light dataset for object detection ... we select those samples in PASCAL VOC with these ten categories and simulate low light corruption to form the training set, dubbed as VOC_dark_train. |
| Dataset Splits | No | The paper mentions training on 'VOC_fog_train' and 'VOC_dark_train' and testing on 'VOC_fog_test', 'RTTS', 'VOC_dark_test', and 'Ex Dark'. It specifies the number of images in these train and test sets but does not explicitly mention a separate validation set or details about its split or usage. |
| Hardware Specification | Yes | All experiments are conducted on an Nvidia Tesla V100 PCIe 32G GPU and implemented by Py Torch. |
| Software Dependencies | No | The paper states that experiments were 'implemented by Py Torch' but does not specify a version number for PyTorch or any other software libraries or dependencies used. |
| Experiment Setup | Yes | We fine-tuned models with batch 1 at each iteration without cropping since the adversarial attack on detectors requires the whole image as input. The tuning number of epochs is 10 for all restoration models. The learning rate is set 1e-4 and the optimizer is ADAM with default settings for all models. ... We set λ as 1e-6. |