Terrace-based Food Counting and Segmentation

Authors: Huu-Thanh Nguyen, Chong-Wah Ngo2364-2372

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experiments are conducted on four datasets: Dimsum, Sushi, Cookie and Mixed dishes, with statistics summarized in Table 2. Each dataset is split into the proportion of 70:20:10 for training, testing and validation respectively. Table 3 compares the performances of counting and instance map segmentation between the terrace model without counting pathway (single-task) and the proposed terrace with a counting included (multi-task). Table 4 lists the performances of different approaches. As noted, multi-task terrace consistently outperforms all the methods across four different datasets.
Researcher Affiliation Academia Huu-Thanh Nguyen,1 Chong-Wah Ngo, 2 1 Department of Computer Science, City University of Hong Kong 2 School of Computing and Information Systems, Singapore Management University tnguyenhu2-c@my.cityu.edu.hk, cwngo@smu.edu.sg
Pseudocode Yes The clustering algorithm is summarized as following: Step 1: Perform the connected component analysis to cluster pixels being labeled as layer k. Instance ids are then assigned to each cluster. The pixels neighbouring to the border of layer k is put in a queue S. Step 2: Set k = k 1 and initialize an empty queue B. 2a: Retrieve a pixel p from the queue S and traverse the neighbours of p. 2b: Propagate the instance id of p to the pixels that are classified to layer k. Add these pixels to queue S. 2c: If any of the neighbours are not classified to kth layer, p is regarded as a pixel located at the border across two layers. Add p to the queue B. 2d: Repeat step-2a until the queue S is empty. Step 3: Copy queue B to queue S. Step 4: Repeat step 2 until k = 1.
Open Source Code No The paper does not provide a direct link to open-source code for the methodology described, nor does it state that the code will be made publicly available.
Open Datasets Yes The last dataset, Mixed dishes, is contributed by (Wang et al. 2019b), where each image is composed of multiple dishes placed on a plate. The images are collected from different canteens in a university. (Wang et al. 2019b) is listed in the references, indicating an established dataset.
Dataset Splits Yes Each dataset is split into the proportion of 70:20:10 for training, testing and validation respectively.
Hardware Specification Yes These deep learning networks are built on Res Net-50 backbone (He et al. 2016) and run on a single GPU of Ge Force GTX 1080.
Software Dependencies No The paper mentions using ResNet-50 as a backbone and Adam optimizer, but does not provide specific version numbers for software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes stochastic gradient descent with warm restarts strategy is employed to adjust the learning rate in the ranges of [10 6, 10 4]. The cycle length is set equal to 32 times higher than the batch size per epoch. All the models are trained with Adam optimizer and the batch size is set to 16. In the experiment, the model training is stopped after 512 epoches when training loss converges. The trade-off parameters in Equation 10 are set to λRC = λCONT OUR = 1.