Pelee: A Real-Time Object Detection System on Mobile Devices
Authors: Robert J. Wang, Xiang Li, Charles X. Ling
NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On Image Net ILSVRC 2012 dataset, our proposed Pelee Net achieves a higher accuracy and over 1.8 times faster speed than Mobile Net and Mobile Net V2 on NVIDIA TX2. Meanwhile, Pelee Net is only 66% of the model size of Mobile Net. We then propose a real-time object detection system by combining Pelee Net with Single Shot Multi Box Detector (SSD) method and optimizing the architecture for fast speed. Our proposed detection system2, named Pelee, achieves 76.4% m AP (mean average precision) on PASCAL VOC2007 and 22.4 m AP on MS COCO dataset at the speed of 23.6 FPS on i Phone 8 and 125 FPS on NVIDIA TX2. |
| Researcher Affiliation | Academia | Robert J. Wang, Xiang Li, Charles X. Ling Department of Computer Science University of Western Ontario London, Ontario, Canada, N6A 3K7 {jwan563,lxiang2,charles.ling}@uwo.ca |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. Figures show architectural diagrams (e.g., 'Structure of 2-way dense layer', 'Structure of stem block') but not pseudocode. |
| Open Source Code | Yes | The code and models are available at: https://github.com/Robert-Jun Wang/Pelee |
| Open Datasets | Yes | On Image Net ILSVRC 2012 dataset, our proposed Pelee Net achieves a higher accuracy... Experimental results on Stanford Dogs [8] dataset... Both training data and validation data are exactly copied from the ILSVRC 2012 dataset. [...] The models are trained on the COCO train+val dataset excluding 5000 minival images and evaluated on the test-dev2015 set. |
| Dataset Splits | Yes | Number of training images: 150,466, Number of validation images: 6,000 [...] The models are trained on the COCO train+val dataset excluding 5000 minival images and evaluated on the test-dev2015 set. |
| Hardware Specification | Yes | This section evaluates the performance of efficient models on i Phone 8 and NVIDIA TX2 embedded platform. [...] Our Pelee Net is trained by Py Torch with mini-batch size 512 on two GPUs. |
| Software Dependencies | No | The paper mentions 'Py Torch', 'Caffe [20]', 'NVIDIA Tensor RT4.0 library', and 'Core ML library'. Only Tensor RT has a version specified (4.0). For a 'Yes' classification, the criteria require multiple key software components with their versions, which is not met here as PyTorch, Caffe, and Core ML lack version numbers. |
| Experiment Setup | Yes | All our models in this section are trained by Py Torch with mini-batch size 256 for 120 epochs. We follow most of the training settings and hyper-parameters used in Res Net on ILSVRC 2012. [...] Our Pelee Net is trained by Py Torch with mini-batch size 512 on two GPUs. The model is trained with a cosine learning rate annealing schedule, similar to what is used by [18] and [19]. The initial learning rate is set to 0.18 and the total amount of epochs is 120. We then fine tune the model with the initial learning rate of 5e-3 for 20 epochs. [...] The batch size is set to 32. The learning rate is set to 0.005 initially, then it decreased by a factor of 10 at 80k and 100k iterations,respectively. The total iterations are 120K. [...] The batch size is set to 128. We first train the model with the learning rate of 10 2 for 70k iterations, and then continue training for 10k iterations with 10 3 and 20k iterations with 10 4. |