{
  "id": 613701,
  "title": "Can behaviors overlap?",
  "url": "/competitions/MABe-mouse-behavior-detection/discussion/613701",
  "author_name": "Ravon",
  "post_date": "2025-10-29T02:28:29.235000",
  "votes": 4,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Can the mouse do multiple actions at the same time? (eg. sniff and approach), but the annotator only annotated one of them? Not sure if we should be using multi-class or multi-label</p>",
  "messages": [
    {
      "id": 3308280,
      "postDate": "2025-10-29T02:28:29.237Z",
      "content": "<p>Can the mouse do multiple actions at the same time? (eg. sniff and approach), but the annotator only annotated one of them? Not sure if we should be using multi-class or multi-label</p>",
      "rawMarkdown": "Can the mouse do multiple actions at the same time? (eg. sniff and approach), but the annotator only annotated one of them? Not sure if we should be using multi-class or multi-label",
      "votes": 4
    },
    {
      "id": 3354078,
      "postDate": "2025-11-30T06:20:57.710Z",
      "content": "<p>Your submission CSV cannot have overlapping actions or it will throw an error. Each frame, agent, target triplet (within one video id) can have a maximum of 1 action.</p>",
      "rawMarkdown": "Your submission CSV cannot have overlapping actions or it will throw an error. Each frame, agent, target triplet (within one video id) can have a maximum of 1 action."
    },
    {
      "id": 3353757,
      "postDate": "2025-11-30T00:19:07.180Z",
      "content": "<h2>Analysis Results: Overlapping Behavior Labels in MABe Train Annotations</h2>\n<p><strong>Yes, overlapping behavior labels ARE allowed for the same agent_id mouse in the MABe dataset.</strong></p>\n<h3>Key Findings</h3>\n<p>I analyzed 6,792 behavior annotations from 57 videos across 16 different labs and found <strong>32 cases of overlapping behavior labels</strong> for the same agent_id mouse. This represents about 0.47% of all behaviors analyzed.</p>\n<h3>Types of Overlapping Behaviors Observed</h3>\n<ol>\n<li><strong>Defensive sequences</strong>: <code>defend</code> behaviors overlapping with <code>freeze</code> behaviors</li>\n<li><strong>Attack sequences</strong>: <code>approach</code> behaviors overlapping with subsequent <code>attack</code> actions  </li>\n<li><strong>Social interactions</strong>: Brief <code>reciprocalsniff</code> bouts overlapping with other ongoing behaviors</li>\n<li><strong>Avoidance patterns</strong>: <code>avoid</code> behaviors overlapping with <code>rear</code> actions</li>\n<li><strong>Complex attack chains</strong>: <code>chaseattack</code> overlapping with <code>attack</code> phases</li>\n</ol>\n<h3>Example Cases</h3>\n<p><strong>Video 1254387004, Agent 2</strong>:</p>\n<ul>\n<li><code>defend</code> (frames 6224-6428) overlaps with previous behavior</li>\n<li><code>freeze</code> (frames 15220-15328) overlaps with previous behavior ending at frame 15332\n<strong>Video 1717182687, Agent 2</strong>:</li>\n<li>Complex sequence with <code>avoid</code> → <code>rear</code> → <code>attack</code> where multiple behaviors overlap temporally</li>\n</ul>\n<h3>Implications for Your Modeling</h3>\n<p>Since overlapping labels are allowed in the ground truth data, your models should be designed to handle:</p>\n<ul>\n<li><strong>Multi-label scenarios</strong> where a single frame/temporal window can have multiple behavior labels</li>\n<li><strong>Temporal dependencies</strong> between overlapping behaviors</li>\n<li><strong>Sequence modeling</strong> that accounts for behavior transitions\nThe competition uses F-beta scoring, which is well-suited for handling multi-label predictions that may result from overlapping behaviors.</li>\n</ul>",
      "rawMarkdown": "\n\n## Analysis Results: Overlapping Behavior Labels in MABe Train Annotations\n\n**Yes, overlapping behavior labels ARE allowed for the same agent_id mouse in the MABe dataset.**\n\n### Key Findings\n\nI analyzed 6,792 behavior annotations from 57 videos across 16 different labs and found **32 cases of overlapping behavior labels** for the same agent_id mouse. This represents about 0.47% of all behaviors analyzed.\n\n### Types of Overlapping Behaviors Observed\n\n1. **Defensive sequences**: `defend` behaviors overlapping with `freeze` behaviors\n2. **Attack sequences**: `approach` behaviors overlapping with subsequent `attack` actions  \n3. **Social interactions**: Brief `reciprocalsniff` bouts overlapping with other ongoing behaviors\n4. **Avoidance patterns**: `avoid` behaviors overlapping with `rear` actions\n5. **Complex attack chains**: `chaseattack` overlapping with `attack` phases\n\n### Example Cases\n\n**Video 1254387004, Agent 2**:\n- `defend` (frames 6224-6428) overlaps with previous behavior\n- `freeze` (frames 15220-15328) overlaps with previous behavior ending at frame 15332\n\n**Video 1717182687, Agent 2**:\n- Complex sequence with `avoid` → `rear` → `attack` where multiple behaviors overlap temporally\n\n### Implications for Your Modeling\n\nSince overlapping labels are allowed in the ground truth data, your models should be designed to handle:\n- **Multi-label scenarios** where a single frame/temporal window can have multiple behavior labels\n- **Temporal dependencies** between overlapping behaviors\n- **Sequence modeling** that accounts for behavior transitions\n\nThe competition uses F-beta scoring, which is well-suited for handling multi-label predictions that may result from overlapping behaviors.\n",
      "replies": [
        {
          "id": 3353760,
          "postDate": "2025-11-30T00:20:34.800Z",
          "content": "<p>ie: less than 0.5% behavior bouts are overlapping. Not a serious concern your model could assume no overlapping and be much simpler.</p>",
          "rawMarkdown": "ie: less than 0.5% behavior bouts are overlapping. Not a serious concern your model could assume no overlapping and be much simpler."
        }
      ]
    },
    {
      "id": 3308369,
      "postDate": "2025-10-29T09:09:23.290Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 3354078,
      "author_name": "Chris Deotte",
      "author_url": "",
      "post_date": "2025-11-30T06:20:57.710000",
      "content": "<p>Your submission CSV cannot have overlapping actions or it will throw an error. Each frame, agent, target triplet (within one video id) can have a maximum of 1 action.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3353757,
      "author_name": "Aaron Toderash",
      "author_url": "",
      "post_date": "2025-11-30T00:19:07.180000",
      "content": "<h2>Analysis Results: Overlapping Behavior Labels in MABe Train Annotations</h2>\n<p><strong>Yes, overlapping behavior labels ARE allowed for the same agent_id mouse in the MABe dataset.</strong></p>\n<h3>Key Findings</h3>\n<p>I analyzed 6,792 behavior annotations from 57 videos across 16 different labs and found <strong>32 cases of overlapping behavior labels</strong> for the same agent_id mouse. This represents about 0.47% of all behaviors analyzed.</p>\n<h3>Types of Overlapping Behaviors Observed</h3>\n<ol>\n<li><strong>Defensive sequences</strong>: <code>defend</code> behaviors overlapping with <code>freeze</code> behaviors</li>\n<li><strong>Attack sequences</strong>: <code>approach</code> behaviors overlapping with subsequent <code>attack</code> actions  </li>\n<li><strong>Social interactions</strong>: Brief <code>reciprocalsniff</code> bouts overlapping with other ongoing behaviors</li>\n<li><strong>Avoidance patterns</strong>: <code>avoid</code> behaviors overlapping with <code>rear</code> actions</li>\n<li><strong>Complex attack chains</strong>: <code>chaseattack</code> overlapping with <code>attack</code> phases</li>\n</ol>\n<h3>Example Cases</h3>\n<p><strong>Video 1254387004, Agent 2</strong>:</p>\n<ul>\n<li><code>defend</code> (frames 6224-6428) overlaps with previous behavior</li>\n<li><code>freeze</code> (frames 15220-15328) overlaps with previous behavior ending at frame 15332\n<strong>Video 1717182687, Agent 2</strong>:</li>\n<li>Complex sequence with <code>avoid</code> → <code>rear</code> → <code>attack</code> where multiple behaviors overlap temporally</li>\n</ul>\n<h3>Implications for Your Modeling</h3>\n<p>Since overlapping labels are allowed in the ground truth data, your models should be designed to handle:</p>\n<ul>\n<li><strong>Multi-label scenarios</strong> where a single frame/temporal window can have multiple behavior labels</li>\n<li><strong>Temporal dependencies</strong> between overlapping behaviors</li>\n<li><strong>Sequence modeling</strong> that accounts for behavior transitions\nThe competition uses F-beta scoring, which is well-suited for handling multi-label predictions that may result from overlapping behaviors.</li>\n</ul>",
      "votes": 0,
      "replies": [
        {
          "id": 3353760,
          "author_name": "Aaron Toderash",
          "author_url": "",
          "post_date": "2025-11-30T00:20:34.800000",
          "content": "<p>ie: less than 0.5% behavior bouts are overlapping. Not a serious concern your model could assume no overlapping and be much simpler.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3308369,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-10-29T09:09:23.290000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3308280": "Can the mouse do multiple actions at the same time? (eg. sniff and approach), but the annotator only annotated one of them? Not sure if we should be using multi-class or multi-label",
    "3354078": "Your submission CSV cannot have overlapping actions or it will throw an error. Each frame, agent, target triplet (within one video id) can have a maximum of 1 action.",
    "3353757": "\n\n## Analysis Results: Overlapping Behavior Labels in MABe Train Annotations\n\n**Yes, overlapping behavior labels ARE allowed for the same agent_id mouse in the MABe dataset.**\n\n### Key Findings\n\nI analyzed 6,792 behavior annotations from 57 videos across 16 different labs and found **32 cases of overlapping behavior labels** for the same agent_id mouse. This represents about 0.47% of all behaviors analyzed.\n\n### Types of Overlapping Behaviors Observed\n\n1. **Defensive sequences**: `defend` behaviors overlapping with `freeze` behaviors\n2. **Attack sequences**: `approach` behaviors overlapping with subsequent `attack` actions  \n3. **Social interactions**: Brief `reciprocalsniff` bouts overlapping with other ongoing behaviors\n4. **Avoidance patterns**: `avoid` behaviors overlapping with `rear` actions\n5. **Complex attack chains**: `chaseattack` overlapping with `attack` phases\n\n### Example Cases\n\n**Video 1254387004, Agent 2**:\n- `defend` (frames 6224-6428) overlaps with previous behavior\n- `freeze` (frames 15220-15328) overlaps with previous behavior ending at frame 15332\n\n**Video 1717182687, Agent 2**:\n- Complex sequence with `avoid` → `rear` → `attack` where multiple behaviors overlap temporally\n\n### Implications for Your Modeling\n\nSince overlapping labels are allowed in the ground truth data, your models should be designed to handle:\n- **Multi-label scenarios** where a single frame/temporal window can have multiple behavior labels\n- **Temporal dependencies** between overlapping behaviors\n- **Sequence modeling** that accounts for behavior transitions\n\nThe competition uses F-beta scoring, which is well-suited for handling multi-label predictions that may result from overlapping behaviors.\n",
    "3308369": ""
  }
}