March 30, 2023
How Does LandingLens Stack Up?
LandingLens, the cloud-based computer vision platform from LandingAI, offers novel techniques to achieve both fast and performance-oriented training out-of-box.
LandingLens training offers a cutting-edge model that can play a crucial role in the data-centric iteration. In fact, a smaller model can act as a good regularizer because it doesn’t have a huge number of parameters to overfit the noise (mislabels) in the dataset. This really helps in identifying marginal or difficult cases.
We decided to put LandingLens to the test by comparing it to open source alternatives. We performed these tests using the default LandingLens training mode, called Fast Mode. We have fine-tuned this training mode to quickly create an accurate model, without requiring you to configure any hyperparameters or settings. The test results show that LandingLens stacks up well against these open source models, which can be difficult to scale and maintain.
The following table shows the results for both LandingLens and the open source alternatives we tested in November 2022. We trained all models using a single NVIDIA A4000 GPU. The number of images used for training is listed next to the dataset link.
As the table shows, LandingLens provides models with fairly good performance out-of-the-box at higher training speeds compared to the alternatives.
This directly helps our users to embrace the data-centric approach and take models to production in a fairly quick time. To create your own computer vision models, sign up for a free trial of LandingLens here.
Object Detection
| Dataset | Test mAP@0.51</span | Training Time (min) | ||
| LandingLens | Open Source Equivalent | LandingLens | Open Source Equivalent | |
| kk_defects (456) | 0.642</span | 0.68 | 6.9 | 16.5 |
| parasite (400) | 0.902</span | 0.92 | 5.1 | 23.2 |
1. The comparison was performed on a Test set. The mAP (Mean Average Precision) score was calculated at a threshold of 0.5 IoU (Intersection over Union). The mAP score measures how precise the model is. A higher score indicates a more accurate model.
2. The performance can be improved by using Advanced Mode training, which offers custom resizing, augmentation, and hyperparameters.
Semantic Segmentation
| Dataset | Test F1 Score1 | Training Time (min) | ||
| LandingLens | Open Source Equivalent | LandingLens | Open Source Equivalent | |
| mvtech_screw (350) | 0.57 | 0.24 | 0.6 | 18.3 |
| mvtech_bottle (220) | 0.78 | 0.68 | 0.4 | 10.5 |
1. The comparison was performed on a Test set. The F1 Score measures precision and recall. A higher score indicates a more accurate model.
Classification
| Dataset | Test F1 Score1 | Training Time (min) | ||
| LandingLens | Open Source Equivalent | LandingLens | Open Source Equivalent | |
| UCI leaf (268) | 1.0 | 0.97 | 3.2 | 7.5 |
| Cracked cement (691) | 0.85 | 0.81 | 8.0 | 19.2 |
1. The comparison was performed on a Test set. The F1 Score measures precision and recall. A higher score indicates a more accurate model.
