Towards Robust Object Detection Invariant to Real-World Domain Shifts
Authors: Qi Fan, Mattia Segu, Yu-Wing Tai, Fisher Yu, Chi-Keung Tang, Bernt Schiele, Dengxin Dai
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the effectiveness of our method for generalizing object detectors under real-world domain shifts. |
| Researcher Affiliation | Academia | 1 The Hong Kong University of Science and Technology, 2 ETH Zurich, 3 Max Planck Institute for Informatics, Saarland Informatics Campus |
| Pseudocode | Yes | B NORMALIZATION PERTURBATION IMPLEMENTATION As shown in the below code block, our Normalization Perturbation (NP) and Normalization Perturbation Plus (NP+) can be easily inserted into popular CNN backbones. import torch import random # Res Net backbone equipped with our Normalization_Perturbation. Class Res Net_backbone(): def __init__(self, ...): def forward(x): p = random.random() x = self.stage_1(x) if p < 0.5 and self.training: # Can be replaced by Normalization_Perturbation_Plus x = Normalization_Perturbation(x) ... def Normalization_Perturbation(feat): ... def Normalization_Perturbation_Plus(feat): |
| Open Source Code | No | While we will make our codes and data public upon acceptance, a competent practitioner can easily reproduce our results. |
| Open Datasets | Yes | The model is trained on the source domain and evaluated on the unseen target domain, i.e., Cityscapes (C) (Cordts et al., 2016) Foggy Cityscapes (F) (Sakaridis et al., 2018), and Sim10k (S) (Johnson-Roberson et al., 2017) Cityscapes (C). We use the Faster R-CNN (Ren et al., 2015) model (with Image Net (Deng et al., 2009)-pretrained Res Net-50 (He et al., 2016) backbone) as the baseline. The detection performance is evaluated using the mean average precision (m AP) metric with the threshold of 0.5. For better comparisons, we also conduct classification DG experiments following the experiment setups in Mixstyle (Zhou et al., 2020b) and DSU (Li et al., 2022), which are closely related methods to our NP method. Refer to the appendix for full experimental details. |
| Dataset Splits | Yes | Specifically, there are two settings: Domain Shift by Time of Day, where the model is trained on BDD100k (Yu et al., 2020) daytime/night train set and evaluated on the night/daytime val set, and Cross-Camera Domain Shift, where the model is trained on Waymo (Sun et al., 2020) Front Left/Right train set and evaluated on BDD100k night val set. |
| Hardware Specification | No | Table 7 mentions "# GPU 4" under the experiment setting details, but it does not specify the particular models of the GPUs used (e.g., NVIDIA A100, Tesla V100). |
| Software Dependencies | No | Table 7 mentions "codebase Detectron2-Cyc Conf Confusion mmdetection" and Appendix B provides Python code snippets, but no specific version numbers for these software components or libraries (e.g., PyTorch 1.9, Detectron2 version) are provided. |
| Experiment Setup | Yes | Table 7: Object detection experiment setting details. batch size 16 LR scheduler Warmup Multi Step LR base LR 0.02 gamma 0.1 momentum 0.9 weight decay 0.0001 warmup method linear warmup iters 1000 warmup factor 1.0 / 1000 LR steps (36000, 48000) (12000, 16000) training iterations 52500 17500 training (min, max) size (800, 1333) testing (min, max) size (800, 1333) |