Exploring Forensic Dental Identification with Deep Learning
Authors: Yuan Liang, Weikun Han, Liang Qiu, Chen Wu, Yiting Shao, Kun Wang, Lei He
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 6 Experimental Results, Table 2: Comparison to the existing dental forensic identification approaches and strong CNN baselines., 6.3 Ablation Studies |
| Researcher Affiliation | Collaboration | Yuan Liang1,2, Weikun Han1, Liang Qiu1, Chen Wu1, Yiting Shao1, Kun Wang1 , Lei He1 1University of California, Los Angeles 2Topaz Labs liangyuandg@ucla.edu |
| Pseudocode | No | The paper describes the methodology in prose and figures, but does not include structured pseudocode or an algorithm block. |
| Open Source Code | Yes | Related data and codes can be found at https://github.com/liangyuandg/Fo ID. |
| Open Datasets | Yes | The training set is derived from an existing DNS Panoramic dataset [49], which contains 543 panoramic radiographs and exist common dental variations of tooth reduction and artifact addition. |
| Dataset Splits | No | The paper describes the training and testing sets, but does not explicitly detail a separate validation set split or its specific proportions/counts. |
| Hardware Specification | Yes | All experiments are conducted on servers with three Ge Force GTX 1080 Ti GPUs. |
| Software Dependencies | No | The paper mentions software components like ResNet-34, Inception-ResNet, and Adam optimizer but does not specify their version numbers or the versions of underlying frameworks like PyTorch or TensorFlow. |
| Experiment Setup | Yes | The output embedding of both encoders has a 512 channels per input attention patch. By following [15], the single layer linear encoder α maps an attention embedding to 128 channels; The transformer holds a hidden embedding dimension of 128 and a multi-attention head number of 4; Meanwhile, the final encoder β consists a fully connected layer and batch normalization, mapping the transform s outputs to 13 channels per attention patch, and thus leading to a dimension of 494 for the final instance embedding. The deep supervision regressor and [class] embedding regressor are both implemented with two fully connected layers, the output feature dimensions of which are 256 and the total count of training instances respectively. Batch normalization and Re LU activation are applied in both regressors. Regarding data pre-processing, all the radiography instances are normalized to a pixel spacing of 0.11mm 0.11mm, and are normalized to a same color space with histogram matching. The attention patch dimensions h and w are set as the minimal embodying size for all anatomies, which are 28.16mm and 21.56mm respectively. Unless otherwise specified, the batch size is set to 6 and the augmentation repeat is set to 5 per instance for all the training. Unless otherwise specified, Adam optimizer is employed with an initial learning rate of 1e-4, and is reduced to a half every 30 epochs. All the methods are trained for 100 epochs. For the online hard pair mining, the positive and negative similarity margins are set to 0.6 and 0.4 respectively. |