config.yml

twinspect: 0.1.0
algorithms:
  - name: ISCC Text-Code V0 64-Bit
    label: text_code_v0_64
    mode: text
    function: twinspect.algos.iscc:text_code_v0_64
    info: The ISCC Text-Code is a similarity preserving hash designed to cluster and match 
      near-duplicate text documents that have undergone format conversion or minor edits.
    url: https://github.com/iscc/iscc-core/blob/main/iscc_core/code_content_text.py
    dependencies:
      - iscc-sdk>=0.8.5
  - name: ISCC Image-Code V0 64-Bit
    label: image_code_v0_64
    mode: image
    function: twinspect.algos.iscc:image_code_v0_64
    info: The ISCC Image-Code is a similarity preserving perceptual hash designed to cluster 
      and match near-duplicate images that have undergone format conversion or minor edits.
    url: https://github.com/iscc/iscc-core/blob/main/iscc_core/code_content_image.py
    dependencies:
      - iscc-sdk>=0.8.5
  - name: ISCC Audio-Code V0 64-Bit
    label: audio_code_v0_64
    mode: audio
    function: twinspect.algos.iscc:audio_code_v0_64
    info: The ISCC Audio-Code is a similarity preserving hash based on chromaprint and designed
      to cluster and match near-duplicate audio files that have undergone format conversion, 
      transcoding, compression and other minor edits.
    url: https://github.com/iscc/iscc-core/blob/main/iscc_core/code_content_audio.py
    dependencies:
      - iscc-sdk>=0.8.5
  - name: ISCC Video-Code V0 64-Bit
    label: video_code_v0_64
    mode: video
    function: twinspect.algos.iscc:video_code_v0_64
    info: The ISCC Video-Code is a similarity preserving hash based on the MPEG-7 Video 
      Signature and is designed to cluster and match near-duplicate videos that have undergone 
      format conversion or minor edits.
    url: https://github.com/iscc/iscc-core/blob/main/iscc_core/code_content_video.py
    dependencies:
      - iscc-sdk>=0.8.5
  - name: ISCC Image-Code-S 64-Bit
    label: image_code_s_64
    mode: image
    function: twinspect.algos.iscc:image_code_s_64
    info: The ISCC Image-Code-S is a semantic image similarity hash based on a deep neural 
      network (DINOv2). It is designed to cluster and match semantically similar images 
      independent of visual transformations.
    url: https://github.com/iscc/iscc-sci
    dependencies:
      - iscc-sci
      - onnxruntime-gpu
  - name: ISCC Image-Code-SC 128-Bit
    label: image_code_sc_128
    mode: image
    function: twinspect.algos.ensemble:placeholder
    info: Combined 128-bit image code concatenating the 64-bit semantic code (iscc-sci/DINOv2) 
      with the 64-bit perceptual code (iscc-sdk). Captures both semantic similarity and visual 
      characteristics.
    url: https://github.com/iscc/twinspect
    dependencies: []
    ensemble_of:
      - image_code_s_64
      - image_code_v0_64
  - name: ISCC Text-Code-S 64-Bit
    label: text_code_s_64
    mode: text
    function: twinspect.algos.iscc:text_code_s_64
    info: The ISCC Text-Code-S is a semantic text similarity hash based on a deep neural 
      network (DeBERTa). It is designed to cluster and match semantically similar text 
      independent of textual transformations, supporting cross-lingual similarity matching.
    url: https://github.com/iscc/iscc-sct
    dependencies:
      - iscc-sct
      - onnxruntime-gpu
  - name: ISCC Text-Code-SC 128-Bit
    label: text_code_sc_128
    mode: text
    function: twinspect.algos.ensemble:placeholder
    info: Combined 128-bit text code concatenating the 64-bit semantic code (iscc-sct/DeBERTa) 
      with the 64-bit perceptual code (iscc-sdk). Captures both semantic similarity and textual
      characteristics.
    url: https://github.com/iscc/twinspect
    dependencies: []
    ensemble_of:
      - text_code_s_64
      - text_code_v0_64
datasets:
  - name: MIRFLICKR-MFND
    label: mirflickr_mfnd
    info: The MFND benchmark ([Connor et al., 
      2015](http://dx.doi.org/10.5220/0005359705650571)) is a subset of the real-world 
      MIRFLICKR dataset ([Huiskes & Lew, 2008](https://doi.org/10.1145/1460096.1460104)) with 
      [annotations](http://www.mir-flickr-near-duplicates.appspot.com/) for near duplicates 
      (IND). The Twinspect benchmark automatically downloads and reproduces the tested dataset.
    url: https://mfnd.similarity.eu/data/truthfiles/polito/IND_clusters.txt
    mode: image
    installer: twinspect.datasets.mfnd:install
  - name: ISCC-FMA-10k
    label: iscc_fma_10k
    info: The ISCC-FMA-10k benchmark is a subset of [Free Music Archive 
      Dataset](https://doi.org/10.48550/arXiv.1612.01840). The subset is generated by 
      collecting 5000 random audio files (longer than 60 seconds). Additionaly 10 synthetic 
      transformations are applied to a random selection 500 of the audio files. The Twinspect 
      benchmark automatically downloads and reproduces the tested dataset.
    url: https://os.unil.cloud.switch.ch/fma/fma_full.zip
    mode: audio
    installer: twinspect.datasets.fma:install
    samples: 5000
    clusters: 500
    seed: 0
  - name: NewsEdits-Reuters-1000
    label: newsedits_reuters_1000
    info: 1000-cluster subset of the [NewsEdits](https://github.com/isi-nlp/NewsEdits) Reuters 
      dataset containing news article revisions (2012-2020). Filtered using LCS similarity 
      (>=70%) and length variation (<=20%) to ensure genuine near-duplicates. Cross-cluster 
      deduplication prevents duplicate articles.
    url: https://github.com/isi-nlp/NewsEdits
    mode: text
    installer: twinspect.datasets.newsedits:install
    clusters: 1000
    min_content_length: 1000
    max_length_variation: 0.20
    min_text_overlap: 0.70
transformations:
  - name: Trim 1 Second
    label: trim-1s-both
    info: Remove 1 seconds of audio from start and end
    mode: audio
    function: twinspect.transformations.audio:trim
    params:
      - 1
      - both
  - name: Trim 5 Seconds
    label: trim-5s-both
    info: Remove 5 seconds of audio from start and end
    mode: audio
    function: twinspect.transformations.audio:trim
    params:
      - 5
      - both
  - name: Fade 8 Seconds
    label: fade-8s-both
    info: Fade in/out 8 seconds at start and end
    mode: audio
    function: twinspect.transformations.audio:fade
    params:
      - 8
      - both
  - name: Transcode MP3 128kbps
    label: transcode-mp3-128kbps
    info: Transcode audio to 128kbps MP3
    mode: audio
    function: twinspect.transformations.audio:transcode
    params:
      - mp3
      - 128
  - name: Transcode OGG 64kbps
    label: transcode-ogg-64kbps
    info: Transcode audio to 64kbps OGG
    mode: audio
    function: twinspect.transformations.audio:transcode
    params:
      - ogg
      - 64
  - name: Transcode AAC 32kbps
    label: transcode-aac-32kbps
    info: Transcode audio to 32kbps AAC
    mode: audio
    function: twinspect.transformations.audio:transcode
    params:
      - aac
      - 32
  - name: Compress
    label: compress-medium
    info: Apply audio compression (attack 10, release 200, ratio 3, threshold -20)
    mode: audio
    function: twinspect.transformations.audio:compress
    params:
      - medium
  - name: Equalize
    label: equalize
    info: Equalize audio (ffmpeg equalizer=f=1000:t=o:w=200:g=10)
    mode: audio
    function: twinspect.transformations.audio:equalize
  - name: Echo
    label: echo
    info: Apply echo effect (ffmpeg aecho=0.8:0.7:60:0.2)
    mode: audio
    function: twinspect.transformations.audio:echo
  - name: Loudness Normalization
    label: loudnorm
    info: Apply loudness normalization (ffmpeg loudnorm=I=-16:TP=-1.5:LRA=11)
    mode: audio
    function: twinspect.transformations.audio:loudnorm
metrics:
  - name: Execution Speed
    label: speed
    function: twinspect.metrics.speed:speed
  - name: Effectiveness (Precission, Recall, F1-Score)
    label: effectiveness
    function: twinspect.metrics.eff:effectiveness
  - name: Robustness against transformations
    label: robustness
    function: twinspect.metrics.robustness:robustness
  - name: Distribution (All-Pairs Hamming Distances)
    label: distribution
    function: twinspect.metrics.distribution:distribution
benchmarks:
  - algorithm_label: image_code_v0_64
    dataset_label: mirflickr_mfnd
    metric_labels:
      - speed
      - effectiveness
      - distribution
    active: true
  - algorithm_label: image_code_s_64
    dataset_label: mirflickr_mfnd
    metric_labels:
      - speed
      - effectiveness
      - distribution
    active: true
  - algorithm_label: image_code_sc_128
    dataset_label: mirflickr_mfnd
    metric_labels:
      - speed
      - effectiveness
      - distribution
    active: true
  - algorithm_label: audio_code_v0_64
    dataset_label: iscc_fma_10k
    metric_labels:
      - speed
      - effectiveness
      - robustness
      - distribution
    active: true
  - algorithm_label: text_code_v0_64
    dataset_label: newsedits_reuters_1000
    metric_labels:
      - speed
      - effectiveness
      - distribution
    active: true
  - algorithm_label: text_code_s_64
    dataset_label: newsedits_reuters_1000
    metric_labels:
      - speed
      - effectiveness
      - distribution
    active: true
  - algorithm_label: text_code_sc_128
    dataset_label: newsedits_reuters_1000
    metric_labels:
      - speed
      - effectiveness
      - distribution
    active: true