Domain Generalised Faster R-CNN
Authors: Karthik Seemakurthy, Charles Fox, Erchan Aptoula, Petra Bosilj
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The proposed approach is implemented as a Domain Generalised Faster R-CNN and evaluated using four object detection datasets which provide domain metadata (GWHD, Cityscapes, BDD100K, Sim10K) where it exhibits a consistent performance improvement over the baselines. All the codes for replicating the results in this paper can be found at https://github.com/karthikiitm87/domain-generalisation.git |
| Researcher Affiliation | Academia | 1Lincoln Institute of Agri-Food Technology, University of Lincoln, United Kingdom. 2School of Computer Science, University of Lincoln, United Kingdom. 3Faculty of Engineering and Natural Sciences (VPALab), Sabanci University, Türkiye. |
| Pseudocode | Yes | Algorithm 1: Training strategy for domain generalised object detection. |
| Open Source Code | Yes | All the codes for replicating the results in this paper can be found at https://github.com/karthikiitm87/domain-generalisation.git |
| Open Datasets | Yes | Datasets. We demonstrate the generalisation ability of our approach on the following four popular multi-source object detection datasets. GWHD (David et al. 2021): ... Cityscapes (C) (Cordts et al. 2016) ... BDD100k (B) (Yu et al. 2020): ... Sim10k (S) (Johnson-Roberson et al. 2017): |
| Dataset Splits | Yes | GWHD (David et al. 2021): ... The training set has 18 domains with a total of 2943 images while the validation set contains samples captured across 8 different sessions with 1424 images and the test set has data from 21 different sessions with a total of 1434 images. Cityscapes (Cordts et al. 2016): It has a total of 2975 training (from 18 cities) and 500 validation images (from 3 cities). BDD100k (B) (Yu et al. 2020): The train, validation, and test splits has 70K, 10K, 20K images, respectively. Sim10k (S) (Johnson-Roberson et al. 2017): There is no official train and validation splits available for this dataset and hence we randomly split it into 8K images as training set and the rest as validation split. |
| Hardware Specification | Yes | The experiments were implemented using the Py Torch deep learning framework and Torchvision Faster R-CNN library on a NVIDIA RTX 3090 GPU with 24GB of GPU memory. |
| Software Dependencies | No | The paper mentions 'Py Torch deep learning framework and Torchvision Faster R-CNN library' but does not specify version numbers for these software components. |
| Experiment Setup | Yes | From empirical observations, the regularisation constants were set to α1 = 1, α2 = 0.1, α3 = 1, α4 = 0.001, and α5 = 0.05. We used early stopping with a patience of 10 epochs. Adam W (weight decay = 0.0005, learning rate = 0.001, batchsize=2) has been used as optimiser while training with GWHD and Stochastic Gradient Descent (SGD) (weight decay = 0.0005, momentum=0.9, learning rate=2 10 3, batchsize=2) has been used for other datasets (Cityscapes, BDD100K, Sim10K). |