Prajna: Towards Recognizing Whatever You Want from Images without Image Labeling
Authors: Xian-Sheng Hua, Jin Li
AAAI 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments The Prajna system has a large number of components. Due to space limit, we will focus on evaluating how image selection and combined recognition improve the overall recognition accuracy. and Table 3. Experiment results. ID Train. Image Sel. Ratio Top 1 Acc. Top 5 Acc. |
| Researcher Affiliation | Industry | Xian-Sheng Hua, Jin Li Microsoft Research, One Microsoft Way, Redmond, WA 98052, USA {xshua; jinl}@microsoft.com |
| Pseudocode | Yes | Algorithm 1: Search based image recognition and Algorithm 2: Combined recognition |
| Open Source Code | No | The paper describes the proposed methods and system architecture but does not provide any links to source code repositories or explicit statements about code release. |
| Open Datasets | Yes | Four entity sets are used for evaluation: Image Net animal subset, Disneyland attractions, Seattle attractions and Aquarium sea creatures. For Image Net subset, we select all the animal related entities from the Image Net 1K dataset 0, which has 163 synsets. and For the other two entity sets, clicked images from Clickture-Full dataset (Hua 2013) is used as the test data. |
| Dataset Splits | Yes | For Image Net dataset, there are 1.3M images in the training set, 50K images in the evaluation set, as well as 150K image in test set for the 1000 synsets. After selecting all the images that are associated with the 163 animal related synsets, we get 211.9K images for training, 8.15K for validation, and 24.45K for testing. |
| Hardware Specification | No | The paper mentions using "GPUs" (in related work context) and a "distributed system with thousands of nodes" and "computer cluster" for their own system, but no specific hardware models (e.g., NVIDIA A100, Intel Xeon, exact memory). |
| Software Dependencies | No | The paper mentions using a "deep neural network," "K-Means," "online SVM," and an "SGD based online SVM learning" but does not provide specific version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | There are indeed quite a few parameters need to set empirically, which is inevitable for a complex real-world end-to-end system. [...] The best parameters obtained on the validate set is 0.15, 0.8, 0.9, 20, 0.01, 3, 7, and 0.9. and Table 2. Parameter sweeping. ID Parameter Range Interval Th1 Cluster based selection classifier [-1.0, 1.0] 0.05 Th2 Intra-entity propagation selection [0.1, 1.0] 0.1 Th3 Inter-entity propagation selection [0.1, 1.0] 0.1 Th4 Online SVM partition number [20, 50] 5 Th5 Learning rate [0.001, 0.1] 10 times Th6 Rounds of learning on each node [1, 10] 1 Th7 Rounds of model averaging [1, 20] 1 Th8 Relative confidence threshold [0.5, 4] 0.1 |