Gradient Harmonized Single-Stage Detector
Authors: Buyu Li, Yu Liu, Xiaogang Wang8577-8584
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Ablation study on MS COCO demonstrates that without laborious hyper-parameter tuning, both GHM-C and GHM-R can bring substantial improvement for single-stage detector. Without any whistles and bells, the proposed model achieves 41.6 m AP on COCO test-dev set which surpass the state-of-the-art method, Focal Loss (FL) + SL1, by 0.8. |
| Researcher Affiliation | Academia | Buyu Li, Yu Liu, Xiaogang Wang {byli, yuliu, xgwang}@ee.cuhk.edu.hk Multimedia Laboratory, The Chinese University of Hong Kong, Hong Kong |
| Pseudocode | No | The paper describes the proposed mechanism and loss functions using mathematical equations and descriptive text, but it does not include a structured pseudocode block or an explicitly labeled algorithm. |
| Open Source Code | Yes | The code1 is released to facilitate future research. 1https://github.com/libuyu/GHM Detection |
| Open Datasets | Yes | We evaluate our approach on the challenging COCO benchmark (Lin et al. 2014). |
| Dataset Splits | Yes | For training, we follow the common used practice (He et al. 2017; Lin et al. 2017b) to divide the 40k validation set into a 35k subset and a 5k subset. The union of the 35k validation subset and the whole 80k training set are used for training together and denoted as trainval35k set. The 5k validation subset is denoted as minival set and our ablation study is performed on it. |
| Hardware Specification | Yes | The experiments are performed on 1080Ti GPUs. |
| Software Dependencies | No | The paper mentions software components and algorithms like RetinaNet, ResNet, FPN, SGD, and various loss functions (smooth L1, cross entropy, Focal Loss), but it does not specify version numbers for any libraries or frameworks (e.g., PyTorch, TensorFlow, specific Python libraries) that would be needed for replication. |
| Experiment Setup | Yes | The input image scale is set as 800 pixel for all experiments. All the models are optimized by the common used SGD algorithm. We train the models on 8 GPUs with 2 images on each GPU so that the effective mini-batch size is 16. All models are trained for 14 epochs with an initial learning rate of 0.01, which is decreased by a factor 0.1 at the 9th epoch and again at the 12th epoch. We also use a weight decay parameter of 0.0001 and a momentum parameter of 0.9. The only data augmentation operation is horizontal image flipping. For the EMA used in gradient density calculation, we use α = 0.75 for all experiments since the results are insensitive to the exact value of α. We adopts µ = 0.02 for ASL1 loss to get comparable results with SL1 loss and obtain a fair baseline for GHM-R loss. |