{
  "id": 679705,
  "title": "Understand the RTM file",
  "url": "/competitions/cvpr-mslr-2026-track-2/discussion/679705",
  "author_name": "",
  "post_date": "2026-03-03T09:20:32.267526400Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi organizers, I have some questions:\n1) Are the RTM .npy files raw magnitudes, or power, are they linear or in dB/log scale?\n2) Please confirm array shape convention (rows = range bins, columns = time frames?) and units (meters per bin, frames/sec) ? \n3) What are the radar timing parameters: chirps per frame, slow‑time sampling rate, and range bin spacing?<br>\n4) What is the mapping of RTM1/2/3 to physical antenna elements?  </p>\n<p>Thanks for any detail provided.</p>",
  "messages": [
    {
      "id": "3416582",
      "postDate": "03/03/2026 09:20:32",
      "content": "<p>Hi organizers, I have some questions:\n1) Are the RTM .npy files raw magnitudes, or power, are they linear or in dB/log scale?\n2) Please confirm array shape convention (rows = range bins, columns = time frames?) and units (meters per bin, frames/sec) ? \n3) What are the radar timing parameters: chirps per frame, slow‑time sampling rate, and range bin spacing?<br>\n4) What is the mapping of RTM1/2/3 to physical antenna elements?  </p>\n<p>Thanks for any detail provided.</p>",
      "rawMarkdown": "Hi organizers, I have some questions:\n1) Are the RTM .npy files raw magnitudes, or power, are they linear or in dB/log scale?\n2) Please confirm array shape convention (rows = range bins, columns = time frames?) and units (meters per bin, frames/sec) ? \n3) What are the radar timing parameters: chirps per frame, slow‑time sampling rate, and range bin spacing?  \n4) What is the mapping of RTM1/2/3 to physical antenna elements?  \n\nThanks for any detail provided.",
      "votes": null
    },
    {
      "id": "3416638",
      "postDate": "03/03/2026 12:20:05",
      "content": "<p>Hi Vu NDLe, thank you for the questions. Here are the relevant details for the released RTM files.</p>\n<p>1 . RTM values</p>\n<ul>\n<li>The RTM .npy files store float32 values</li>\n<li>They are not raw ADC data, not complex FFT coefficients, and not linear power</li>\n<li>Each RTM is built from the Infineon XENSIV Range Spectrum pipeline applied to the BGT60TR13C data: mean removal enabled, Blackman-Harris windowing on fast time, FFT, magnitude (abs), thresholding at 1e-6, then scaling</li>\n<li>The stored values are in dB, using 20*log10 amplitude scaling (IFX_SCALE_TYPE_DECIBEL_20LOG)</li>\n<li>So the released RTMs represent magnitude in dB (amplitude dB), not power dB</li>\n<li>Please do not apply log scaling again</li>\n</ul>\n<p>2 . Array shape convention and axes</p>\n<ul>\n<li>Each .npy file contains a 2D array with shape [T, R]</li>\n<li>T = number of time frames in the sample</li>\n<li>R = number of stored positive range bins</li>\n<li>Axis convention is:</li>\n<li>- rows -&gt; time frames (slow-time, in chronological order)</li>\n<li>- columns -&gt; range bins (increasing radial distance)</li>\n<li>In other words, the RTM is a time-by-range map, with time on axis 0 and range on axis 1</li>\n</ul>\n<p>3 . Units and timing parameters</p>\n<ul>\n<li>The radar sensor is always the Infineon BGT60TR13C</li>\n<li>The acquisition was performed with a single fixed configuration across the dataset</li>\n<li>Frame repetition time: 0.077041544 s</li>\n<li>This corresponds to a slow-time sampling rate of about 12.98 frames/s</li>\n<li>Sample rate: 1,000,000 Hz</li>\n<li>Center frequency: 60.500000768 GHz</li>\n<li>Configured range resolution: 0.0312315039 m per bin, i.e. about 3.12 cm/bin</li>\n<li>Configured maximum range: 1.6 m</li>\n<li>Therefore, the range axis spans approximately 0 -&gt; 1.6 m in steps of about 3.12 cm</li>\n<li>The reported speed settings are:</li>\n<li>- max speed: 4.10804987 m/s</li>\n<li>- speed resolution: 0.03209735 m/s</li>\n<li>Under this fixed configuration, these values correspond to 256 chirps per frame, which is the same chirp count used in the RS/RDM processing chain</li>\n</ul>\n<p>4 . RTM1/RTM2/RTM3 mapping</p>\n<ul>\n<li>rx_mask = 7 means that all three receive channels are enabled</li>\n<li>tx_mask = 1 means a single transmit channel is used</li>\n<li>Accordingly:</li>\n<li>- RTM1 -&gt; receive channel 1</li>\n<li>- RTM2 -&gt; receive channel 2</li>\n<li>- RTM3 -&gt; receive channel 3</li>\n<li>These are three synchronized views of the same gesture sample, computed independently from the three RX streams</li>\n</ul>\n<p>5 . Additional processing notes</p>\n<ul>\n<li>Only the positive range half is stored in the released RTMs</li>\n<li>The range FFT uses:</li>\n<li>- IFX_FFT_TYPE_R2C</li>\n<li>- mean removal enabled</li>\n<li>- normalized Blackman-Harris window</li>\n<li>The threshold floor is 1e-6, which corresponds to a floor near -120 dB before storage</li>\n</ul>\n<p>So, in summary: the released RTMs are 2D float32 arrays of shape [time_frames, positive_range_bins], with values equal to range-spectrum magnitude in dB, sampled at about 12.98 Hz along time, and with approximately 3.12 cm spacing along the range axis.</p>\n<p>This is an anticipation of our paper (that we will release soon to have all information to write your paper) with baseline for this challenge datas.</p>",
      "rawMarkdown": "Hi Vu NDLe, thank you for the questions. Here are the relevant details for the released RTM files.\n\n1 . RTM values\n- The RTM .npy files store float32 values\n- They are not raw ADC data, not complex FFT coefficients, and not linear power\n- Each RTM is built from the Infineon XENSIV Range Spectrum pipeline applied to the BGT60TR13C data: mean removal enabled, Blackman-Harris windowing on fast time, FFT, magnitude (abs), thresholding at 1e-6, then scaling\n- The stored values are in dB, using 20*log10 amplitude scaling (IFX_SCALE_TYPE_DECIBEL_20LOG)\n- So the released RTMs represent magnitude in dB (amplitude dB), not power dB\n- Please do not apply log scaling again\n\n2 . Array shape convention and axes\n- Each .npy file contains a 2D array with shape [T, R]\n- T = number of time frames in the sample\n- R = number of stored positive range bins\n- Axis convention is:\n- - rows -> time frames (slow-time, in chronological order)\n- - columns -> range bins (increasing radial distance)\n- In other words, the RTM is a time-by-range map, with time on axis 0 and range on axis 1\n\n3 . Units and timing parameters\n- The radar sensor is always the Infineon BGT60TR13C\n- The acquisition was performed with a single fixed configuration across the dataset\n- Frame repetition time: 0.077041544 s\n- This corresponds to a slow-time sampling rate of about 12.98 frames/s\n- Sample rate: 1,000,000 Hz\n- Center frequency: 60.500000768 GHz\n- Configured range resolution: 0.0312315039 m per bin, i.e. about 3.12 cm/bin\n- Configured maximum range: 1.6 m\n- Therefore, the range axis spans approximately 0 -> 1.6 m in steps of about 3.12 cm\n- The reported speed settings are:\n- - max speed: 4.10804987 m/s\n- - speed resolution: 0.03209735 m/s\n- Under this fixed configuration, these values correspond to 256 chirps per frame, which is the same chirp count used in the RS/RDM processing chain\n\n4 . RTM1/RTM2/RTM3 mapping\n- rx_mask = 7 means that all three receive channels are enabled\n- tx_mask = 1 means a single transmit channel is used\n- Accordingly:\n- - RTM1 -> receive channel 1\n- - RTM2 -> receive channel 2\n- - RTM3 -> receive channel 3\n- These are three synchronized views of the same gesture sample, computed independently from the three RX streams\n\n5 . Additional processing notes\n- Only the positive range half is stored in the released RTMs\n- The range FFT uses:\n- - IFX_FFT_TYPE_R2C\n- - mean removal enabled\n- - normalized Blackman-Harris window\n- The threshold floor is 1e-6, which corresponds to a floor near -120 dB before storage\n\nSo, in summary: the released RTMs are 2D float32 arrays of shape [time_frames, positive_range_bins], with values equal to range-spectrum magnitude in dB, sampled at about 12.98 Hz along time, and with approximately 3.12 cm spacing along the range axis.\n\nThis is an anticipation of our paper (that we will release soon to have all information to write your paper) with baseline for this challenge datas.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3416638,
      "author_name": "raffaelemineo",
      "author_url": "",
      "post_date": "03/03/2026 12:20:05",
      "content": "<p>Hi Vu NDLe, thank you for the questions. Here are the relevant details for the released RTM files.</p>\n<p>1 . RTM values</p>\n<ul>\n<li>The RTM .npy files store float32 values</li>\n<li>They are not raw ADC data, not complex FFT coefficients, and not linear power</li>\n<li>Each RTM is built from the Infineon XENSIV Range Spectrum pipeline applied to the BGT60TR13C data: mean removal enabled, Blackman-Harris windowing on fast time, FFT, magnitude (abs), thresholding at 1e-6, then scaling</li>\n<li>The stored values are in dB, using 20*log10 amplitude scaling (IFX_SCALE_TYPE_DECIBEL_20LOG)</li>\n<li>So the released RTMs represent magnitude in dB (amplitude dB), not power dB</li>\n<li>Please do not apply log scaling again</li>\n</ul>\n<p>2 . Array shape convention and axes</p>\n<ul>\n<li>Each .npy file contains a 2D array with shape [T, R]</li>\n<li>T = number of time frames in the sample</li>\n<li>R = number of stored positive range bins</li>\n<li>Axis convention is:</li>\n<li>- rows -&gt; time frames (slow-time, in chronological order)</li>\n<li>- columns -&gt; range bins (increasing radial distance)</li>\n<li>In other words, the RTM is a time-by-range map, with time on axis 0 and range on axis 1</li>\n</ul>\n<p>3 . Units and timing parameters</p>\n<ul>\n<li>The radar sensor is always the Infineon BGT60TR13C</li>\n<li>The acquisition was performed with a single fixed configuration across the dataset</li>\n<li>Frame repetition time: 0.077041544 s</li>\n<li>This corresponds to a slow-time sampling rate of about 12.98 frames/s</li>\n<li>Sample rate: 1,000,000 Hz</li>\n<li>Center frequency: 60.500000768 GHz</li>\n<li>Configured range resolution: 0.0312315039 m per bin, i.e. about 3.12 cm/bin</li>\n<li>Configured maximum range: 1.6 m</li>\n<li>Therefore, the range axis spans approximately 0 -&gt; 1.6 m in steps of about 3.12 cm</li>\n<li>The reported speed settings are:</li>\n<li>- max speed: 4.10804987 m/s</li>\n<li>- speed resolution: 0.03209735 m/s</li>\n<li>Under this fixed configuration, these values correspond to 256 chirps per frame, which is the same chirp count used in the RS/RDM processing chain</li>\n</ul>\n<p>4 . RTM1/RTM2/RTM3 mapping</p>\n<ul>\n<li>rx_mask = 7 means that all three receive channels are enabled</li>\n<li>tx_mask = 1 means a single transmit channel is used</li>\n<li>Accordingly:</li>\n<li>- RTM1 -&gt; receive channel 1</li>\n<li>- RTM2 -&gt; receive channel 2</li>\n<li>- RTM3 -&gt; receive channel 3</li>\n<li>These are three synchronized views of the same gesture sample, computed independently from the three RX streams</li>\n</ul>\n<p>5 . Additional processing notes</p>\n<ul>\n<li>Only the positive range half is stored in the released RTMs</li>\n<li>The range FFT uses:</li>\n<li>- IFX_FFT_TYPE_R2C</li>\n<li>- mean removal enabled</li>\n<li>- normalized Blackman-Harris window</li>\n<li>The threshold floor is 1e-6, which corresponds to a floor near -120 dB before storage</li>\n</ul>\n<p>So, in summary: the released RTMs are 2D float32 arrays of shape [time_frames, positive_range_bins], with values equal to range-spectrum magnitude in dB, sampled at about 12.98 Hz along time, and with approximately 3.12 cm spacing along the range axis.</p>\n<p>This is an anticipation of our paper (that we will release soon to have all information to write your paper) with baseline for this challenge datas.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3416582": "Hi organizers, I have some questions:\n1) Are the RTM .npy files raw magnitudes, or power, are they linear or in dB/log scale?\n2) Please confirm array shape convention (rows = range bins, columns = time frames?) and units (meters per bin, frames/sec) ? \n3) What are the radar timing parameters: chirps per frame, slow‑time sampling rate, and range bin spacing?  \n4) What is the mapping of RTM1/2/3 to physical antenna elements?  \n\nThanks for any detail provided.",
    "3416638": "Hi Vu NDLe, thank you for the questions. Here are the relevant details for the released RTM files.\n\n1 . RTM values\n- The RTM .npy files store float32 values\n- They are not raw ADC data, not complex FFT coefficients, and not linear power\n- Each RTM is built from the Infineon XENSIV Range Spectrum pipeline applied to the BGT60TR13C data: mean removal enabled, Blackman-Harris windowing on fast time, FFT, magnitude (abs), thresholding at 1e-6, then scaling\n- The stored values are in dB, using 20*log10 amplitude scaling (IFX_SCALE_TYPE_DECIBEL_20LOG)\n- So the released RTMs represent magnitude in dB (amplitude dB), not power dB\n- Please do not apply log scaling again\n\n2 . Array shape convention and axes\n- Each .npy file contains a 2D array with shape [T, R]\n- T = number of time frames in the sample\n- R = number of stored positive range bins\n- Axis convention is:\n- - rows -> time frames (slow-time, in chronological order)\n- - columns -> range bins (increasing radial distance)\n- In other words, the RTM is a time-by-range map, with time on axis 0 and range on axis 1\n\n3 . Units and timing parameters\n- The radar sensor is always the Infineon BGT60TR13C\n- The acquisition was performed with a single fixed configuration across the dataset\n- Frame repetition time: 0.077041544 s\n- This corresponds to a slow-time sampling rate of about 12.98 frames/s\n- Sample rate: 1,000,000 Hz\n- Center frequency: 60.500000768 GHz\n- Configured range resolution: 0.0312315039 m per bin, i.e. about 3.12 cm/bin\n- Configured maximum range: 1.6 m\n- Therefore, the range axis spans approximately 0 -> 1.6 m in steps of about 3.12 cm\n- The reported speed settings are:\n- - max speed: 4.10804987 m/s\n- - speed resolution: 0.03209735 m/s\n- Under this fixed configuration, these values correspond to 256 chirps per frame, which is the same chirp count used in the RS/RDM processing chain\n\n4 . RTM1/RTM2/RTM3 mapping\n- rx_mask = 7 means that all three receive channels are enabled\n- tx_mask = 1 means a single transmit channel is used\n- Accordingly:\n- - RTM1 -> receive channel 1\n- - RTM2 -> receive channel 2\n- - RTM3 -> receive channel 3\n- These are three synchronized views of the same gesture sample, computed independently from the three RX streams\n\n5 . Additional processing notes\n- Only the positive range half is stored in the released RTMs\n- The range FFT uses:\n- - IFX_FFT_TYPE_R2C\n- - mean removal enabled\n- - normalized Blackman-Harris window\n- The threshold floor is 1e-6, which corresponds to a floor near -120 dB before storage\n\nSo, in summary: the released RTMs are 2D float32 arrays of shape [time_frames, positive_range_bins], with values equal to range-spectrum magnitude in dB, sampled at about 12.98 Hz along time, and with approximately 3.12 cm spacing along the range axis.\n\nThis is an anticipation of our paper (that we will release soon to have all information to write your paper) with baseline for this challenge datas."
  },
  "source": "meta"
}