jakubmicorek
commented
1 month ago Hello,
thank you for your interest in our research.
The code provided was intended to release the anomaly detection method where the same training and evaluation principles apply to the video features. Unfortunately, due to the extensive maintenance and refactoring required, I am currently unable to provide the complete dataset-specific training and evaluation scripts at this time. I may add the code in the future.
For the object-centric evaluation please find the details in section 2.2 of the supplementary material and have a look at the ablations in Table 3 and Table 4. The code provided by AccI-VAD does not handle the 1:1 mapping of pose with velocity and deep features. This means, that for each bounding box in a frame, the velocity and deep features align; however, the pose features within a frame do not have a direct mapping to the bounding box. For this reason, we decided to follow AccI-VAD's approach and fit/train separate models for each feature type. For training and evaluating we create a TensorDataset with one specific feature type and add the video name and the frame number such that the features can be back-referenced. During training, we additionally shuffle the TensorDataset. To obtain an anomaly score for a single frame, the maximum anomaly score per feature type within a frame is taken and aggregated. Note, as described in section 2.2 of the supplementary material, we standardize the anomaly score by the statistics of the training data by their respective feature type and clip negative values before adding up to obtain the final frame-level anomaly score. For each video, the aggregated anomaly scores are temporally smoothed with a 1d-Gaussian filter. We tested temporal smoothing of 1-7 frames. Finally, the micro and macro AUC-ROC scores are computed.
Please feel free to reach out if you need further assistance.
syrobonkers
Haifu-Ye
jpainam
Hello, Thank you for your excellent research and for making the code available as open-source software. I have tried to reproduce the experiment with your code and noticed that only the Toy Dataset is provided. Could you please provide guidance on how to train and evaluate the model on the Ped2, Avenue, and ShanghaiTech datasets mentioned in your paper?
I understand that for object-level tasks, I should use the pre-extracted features available in the repository. However, I am unsure how to utilize the three types of features—Pose, Deep, and Velocity. Could you explain how these features should be used? Additionally, could you provide information on how to perform the evaluation? If possible, it would be helpful if you could also add the training and evaluation code for these datasets in order to fully reproduce the experiment.
Thank you very much for your assistance.