# BiasBench Dataset BiasBench is a large-scale dataset collection consisting of three sub-datasets capturing event-based data under various hardware bias configurations for the Prophesee EVK4 Camera. Each sub-dataset contains training and validation sets, organized into subgroups based on `bias_diff_on` and `bias_diff_off` thresholds. --- ## Dataset Structure ``` files/ ├── led/ │ ├── data_set/ │ │ ├── group_0_0/ │ │ ├── ... │ │ └── group_5_5/ │ └── validation_set/ │ ├── group_0_0/ │ ├── ... │ └── group_5_5/ │ ├── robot_arm/ │ ├── left_right/ │ │ ├── group_0_0/ │ │ ├── ... │ │ └── group_2_2/ │ └── triangle/ │ ├── group_0_0/ │ ├── ... │ └── group_2_2/ │ └── spinning_dot/ ├── black_dot/ │ ├── group_0_0/ │ ├── ... │ └── group_8_8/ └── grey_dot/ ├── group_0_0/ ├── ... └── group_8_8/ ``` --- ## Bias Configurations Each sub-dataset explores a range of bias values affecting event generation: ### LED - `bias_diff_on`: 120, 100, 80, 60, 40, 20, 0, -20, -40, -60, -80 - `bias_diff_off`: 170, 150, 130, 110, 90, 70, 50, 30, 10, -10, -30 - `bias_fo`: 48, 37, 26, 15, 4, -7, -18, -29 - `bias_hpf`: 116, 100, 84, 68, 52, 36, 20, 4 - `bias_refr`: 225, 145, 65, -15 ### Robot Arm - `bias_diff_on`: 140, 126 112, 99 ,85, 72, 58, 45, 31, 17, 4, -9, -22 -36, -50 - `bias_diff_off`: 190, 175, 161, 147, 132, 118, 104, 90, 75, 61, 47, 32, 18, 4, -10 - `bias_fo`: 55, 33, 12, -8, -30 - `bias_hpf`: 120, 96, 72, 48, 24, 0 - `bias_refr`: 0 ### Spinning Dot - `bias_diff_on`: 140, 130, 120, 110, 100, 90, 80, 70, 60, 50, 40, 30, 20, 10, 0, -10, -20, -30 - `bias_diff_off`: 190, 178, 166, 154, 142, 131, 119, 107, 94, 84, 72, 60, 48, 37, 25, 13, 1, -10 - `bias_fo`: 55, 32, 10, -12, -35 - `bias_hpf`: 120, 96, 72, 48, 24, 0 - `bias_refr`: 200, 126, 53, -20 --- ## File Counts & Sizes ### LED | Group # | bias_diff_on | bias_diff_off | # Files | |---------|------------------|------------------------|---------| | 0 | 120, 100 | 170, 150 | 1024 | | 1 | 80, 60 | 130, 110 | 1024 | | 2 | 40, 20 | 90, 70 | 1024 | | 3 | 0, -20 | 50, 30 | 1024 | | 4 | -40, -60 | 10, -10 | 1024 | **Edge regions:** | Group # | bias_diff_on | bias_diff_off | # Files | |---------|-------------------------|-----------------------------|---------| | 0_5 | 120, 100, 80, 60 | -30 | 1024 | | 1_5 | 40, 20, 0, -20 | -30 | 1024 | | 5_0 | -80 | 170, 150, 130, 110 | 1024 | | 5_1 | -80 | 90, 70, 50, 30 | 1024 | | 5_5 | -40, -60 | -30 | 1024 | | 5_5 | -80 | 10, -10, -30 | 1280 | - **data_set**: 30,976 files (≈ 5.6 TB) - **validation_set**: 30,976 files (≈ 4.8 TB) --- ### Robot Arm | Group # | bias_diff_on | bias_diff_off | # Files | |---------|-------------------------|-----------------------------|---------| | 0 | 140, 126, 112, 99, 85 | 190, 175, 161, 147, 132 | 750 | | 1 | 72, 58, 45, 31, 17 | 118, 104, 90, 75, 61 | 750 | | 2 | 4, -9, -22, -36, -50 | 47, 32, 18, 4, -10 | 750 | - **left_right**: 6,750 files (≈ 830 GB) - **triangle**: 6,750 files (≈ 714 GB) --- ### Spinning Dot | Group # | bias_diff_on | bias_diff_off | # Files | |---------|------------------|------------------------|---------| | 0 | 140, 130 | 190, 178 | 480 | | 1 | 120, 110 | 166, 154 | 480 | | 2 | 100, 90 | 142, 131 | 480 | | 3 | 80, 70 | 119, 107 | 480 | | 4 | 60, 50 | 94, 84 | 480 | | 5 | 40, 30 | 72, 60 | 480 | | 6 | 20, 10 | 48, 37 | 480 | | 7 | 0, -10 | 25, 13 | 480 | | 8 | -20, -30 | 1, -10 | 480 | - **black_dot**: 38,880 files (≈ 380 GB) - **grey_dot**: 38,880 files (≈ 821 GB) --- ## Internal File Sturcture Each recording is saved as an HDF5 file with the following structure: ### File Format: `bias_diff_on` _ `bias_diff_off` _ `bias_fo` _ `bias_hpf` _ `bias_refr` .h5 ### File Contents (HDF5 Tree) ``` .h5 ├── events/ │ ├── x # x-coordinates of events │ ├── y # y-coordinates of events │ ├── p # polarities of events │ └── t # timestamps of events ├── ms_to_idx # index mapping: first event after x milliseconds ├── bias # list of 5 bias values └── attributes ├── width # width of the recording (e.g., 1280) ├── height # height of the recording (e.g., 720) └── sensor # sensor model used (e.g., IMX636) ``` ### Descriptions `events/x, y, p, t`: Arrays of equal length containing the event stream data. May be empty if no events were generated. `ms_to_idx`: A lookup table where ms_to_idx[x] gives the index of the first event with t[i] ≥ x. `bias`: A list of five integers indicating the biases used for this recording, in the same order as encoded in the filename: `bias_diff_on`, `bias_diff_off`, `bias_fo`, `bias_hpf`, `bias_refr` HDF5 attributes: `width`: Width of the sensor image (e.g., 1280) `height`: Height of the sensor image (e.g., 720) `sensor`: String name of the sensor used (e.g., IMX636) ## Download Instructions Due to the dataset's size (≈13.2 TB), it is distributed via a **torrent file**, allowing selective download of specific sub-datasets or bias configurations. ### Steps to Download: 1. **Download the torrent file:** Get the `.torrent` file from the biasbench-2025 folder. 2. **Open the torrent file in a torrent client** 3. **Select the data you need:** - When prompted, choose only the folders of interest (e.g., `robot_arm/left_right/group_1_0/`). - This is useful if you want to download only a specific: - Bias configuration range - Dataset split (e.g., training vs. validation) - Subgroup or scene type 4. **Start the download** The client will only fetch the selected files, reducing the required bandwidth and storage. --- ## Contact For questions, bug reports, or collaboration inquiries, please contact **[Your Name or Organization]** at **[your.email@example.com]**. --- ## License Include license information here (e.g., CC BY 4.0) If you use our dataset, please cite with: ``` @INPROCEEDINGS{ziegler2025biasbenchreproduciblebenchmarktuning, title={BiasBench: A reproducible benchmark for tuning the biases of event cameras}, author={Andreas Ziegler and David Joseph and Thomas Gossard and Emil Moldovan and Andreas Zell}, year={2025}, booktitle = { 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) }, publisher = {IEEE Computer Society}, } ``` ---