Load the data
","metadata":{}},{"cell_type":"markdown","source":"**Load train csv with files description:**","metadata":{}},{"cell_type":"code","source":"dir <- \"/kaggle/input/MABe-mouse-behavior-detection/\"\ntrain_meta <- fread(file.path(dir, \"train.csv\"))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:32:39.629789Z","iopub.execute_input":"2025-10-03T20:32:39.631268Z","iopub.status.idle":"2025-10-03T20:32:39.726785Z","shell.execute_reply":"2025-10-03T20:32:39.724933Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"**Calculate summary per each file for all parquet files in the train_tracking folder:**\n- min/max x and y coordinates\n- first and last frame\n- number of distinct frames\n- total number of frames\n- number of mice\n- number of body parts","metadata":{}},{"cell_type":"code","source":"con <- dbConnect(duckdb())\ndbExecute(con, \"PRAGMA threads=4;\")\n\nbase <- file.path(dir, \"train_tracking\")\n\nquery <- sprintf(\"\nSELECT\n filename,\n MIN(video_frame) AS first_frame,\n MAX(video_frame) AS last_frame,\n COUNT(DISTINCT video_frame) AS num_distinct_frames,\n COUNT(DISTINCT mouse_id) AS num_mice,\n COUNT(DISTINCT bodypart) AS num_bodyparts,\n MIN(x) AS min_x, MAX(x) AS max_x,\n MIN(y) AS min_y, MAX(y) AS max_y\nFROM read_parquet('%s/*/*.parquet', filename=true)\nGROUP BY filename\n\", base)\n\ntic()\ntrack_summary <- as.data.table(dbGetQuery(con, query))\ntoc()\ntrack_summary[, video_id := sub(\"\\\\.parquet$\", \"\", basename(filename))]\ntrack_summary[, filename := NULL]\ntrack_summary[, span_frames := last_frame - first_frame + 1L]\nsetcolorder(track_summary, \"video_id\", before = \"first_frame\")\n\ncat(nrow(track_summary), \"rows,\", ncol(track_summary), \"columns\\n\")\nhead(track_summary)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:32:39.729510Z","iopub.execute_input":"2025-10-03T20:32:39.731011Z","iopub.status.idle":"2025-10-03T20:34:41.235650Z","shell.execute_reply":"2025-10-03T20:34:41.233094Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"Load train annotations","metadata":{}},{"cell_type":"code","source":"ann_files <- dir_ls(\n file.path(dir, \"train_annotation\"),\n recurse = TRUE, glob = \"*.parquet\"\n)\ntrain_ann <- rbindlist(\n lapply(ann_files, \\(f) {\n dt <- as.data.table(read_parquet(f))\n fn <- gsub(\"\\\\.parquet\", \"\", basename(f))\n dt[, video_file := fn]\n }),\n use.names = TRUE, fill = TRUE\n)\nsetnames(train_ann, \"video_file\",\"video_id\")\ncat(\"Number of files:\", length(ann_files))\nhead(train_ann)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:41.239744Z","iopub.execute_input":"2025-10-03T20:34:41.241528Z","iopub.status.idle":"2025-10-03T20:34:50.467641Z","shell.execute_reply":"2025-10-03T20:34:50.465722Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#Missing annotations
","metadata":{}},{"cell_type":"markdown","source":"Сheck whether all videos listed in train.csv have corresponding annotation files and tracking files and vice versa.","metadata":{}},{"cell_type":"code","source":"ann_files_dt <- data.table(file = ann_files)[, .(\n lab_id = basename(dirname(file)),\n video_id = basename(gsub(\".parquet\", \"\", file))\n )\n]\ntrack_files <- dir_ls(\n file.path(dir, \"train_tracking\"),\n recurse = TRUE, glob = \"*.parquet\"\n)\ntrack_files_dt <- data.table(file = track_files)[, .(\n lab_id = basename(dirname(file)),\n video_id = basename(gsub(\".parquet\", \"\", file))\n )\n]\nmeta_files_dt <- train_meta[, .(video_id = as.character(video_id), lab_id)]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:50.470431Z","iopub.execute_input":"2025-10-03T20:34:50.471899Z","iopub.status.idle":"2025-10-03T20:34:50.609063Z","shell.execute_reply":"2025-10-03T20:34:50.607130Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"Calculate number of videos with metadata in train.csv, annotations and tracking:","metadata":{}},{"cell_type":"code","source":"meta_files_dt[, num_in_csv := uniqueN(video_id), by = lab_id]\nmeta_files_dt[, `:=` (\n has_ann = video_id %in% ann_files_dt$video_id,\n has_tracking = video_id %in% track_files_dt$video_id\n)]\nmeta_files_dt[, `:=` (\n num_annotated = sum(has_ann),\n num_tracking = sum(has_tracking)\n), by = lab_id]\nsummary <- meta_files_dt[, .(lab_id, num_in_csv, num_tracking, num_annotated)][\n order(-num_in_csv), unique(.SD)]\nsummary","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:50.611821Z","iopub.execute_input":"2025-10-03T20:34:50.613317Z","iopub.status.idle":"2025-10-03T20:34:50.665324Z","shell.execute_reply":"2025-10-03T20:34:50.663423Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"summary[num_in_csv != num_tracking | num_in_csv != num_annotated]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:50.667961Z","iopub.execute_input":"2025-10-03T20:34:50.669436Z","iopub.status.idle":"2025-10-03T20:34:50.691832Z","shell.execute_reply":"2025-10-03T20:34:50.690054Z"},"_kg_hide-input":false},"outputs":[],"execution_count":null},{"cell_type":"code","source":"no_ann <- meta_files_dt[has_ann == FALSE]\nno_ann[, .(\n miss_videos = uniqueN(video_id),\n `%` = round(uniqueN(video_id) / num_in_csv[1] * 100, 1)\n), by = lab_id]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:50.694486Z","iopub.execute_input":"2025-10-03T20:34:50.696038Z","iopub.status.idle":"2025-10-03T20:34:50.736100Z","shell.execute_reply":"2025-10-03T20:34:50.734243Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"Missing frames
","metadata":{}},{"cell_type":"markdown","source":"Check for gaps in the tracking frame sequence, to detect potential frame drops or incomplete tracking data.","metadata":{}},{"cell_type":"code","source":"check_frames <- track_summary[\n , .(video_id, first_frame, last_frame, span_frames, num_distinct_frames)\n]\ncheck_frames[meta_files_dt, on = .(video_id), lab_id := i.lab_id]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:50.738797Z","iopub.execute_input":"2025-10-03T20:34:50.740285Z","iopub.status.idle":"2025-10-03T20:34:50.778513Z","shell.execute_reply":"2025-10-03T20:34:50.775908Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"First I compare if calculated `span_frames` (last_frame - first_frame + 1) is equal to number of distince frames - meaning there are no gaps:","metadata":{}},{"cell_type":"code","source":"check_frames[, miss_frames := span_frames - num_distinct_frames]\ncheck_frames[, miss_frames_pct := round(miss_frames / span_frames * 100, 1)]\nissues <- check_frames[abs(miss_frames) > 0]\n\nreport_issues(issues, \"Videos with time gaps:\")\nhead(issues[order(-abs(miss_frames_pct))], 10)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:50.781390Z","iopub.execute_input":"2025-10-03T20:34:50.782965Z","iopub.status.idle":"2025-10-03T20:34:50.828657Z","shell.execute_reply":"2025-10-03T20:34:50.826681Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"quantile(issues$miss_frames_pct)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:50.831425Z","iopub.execute_input":"2025-10-03T20:34:50.833015Z","iopub.status.idle":"2025-10-03T20:34:50.850558Z","shell.execute_reply":"2025-10-03T20:34:50.848736Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"Number of frames missmatch
","metadata":{}},{"cell_type":"markdown","source":"Compare the number of frames derived from train.csv (fps × duration) with the span observed in tracking files, to detect inconsistencies in metadata versus actual tracking.","metadata":{}},{"cell_type":"code","source":"check_frames <- track_summary[\n , .(video_id, first_frame, last_frame, span_frames, num_distinct_frames)\n]\ncheck_frames[meta_files_dt, on = .(video_id), lab_id := i.lab_id]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:58.648849Z","iopub.execute_input":"2025-10-03T20:34:58.650407Z","iopub.status.idle":"2025-10-03T20:34:58.680426Z","shell.execute_reply":"2025-10-03T20:34:58.678494Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"**Add video characteristics from train.csv**\n- frames_per_second as `fps`\n- video_duration_sec\n- total frames (frames_per_second * video_duration_sec) as `csv_total_frames` (number of frames in video according to train.csv metadata)","metadata":{}},{"cell_type":"code","source":"train_info <- train_meta[, .(\n video_id = as.character(video_id),\n fps = frames_per_second,\n video_duration_sec = video_duration_sec,\n csv_total_frames = round(frames_per_second * video_duration_sec)\n)]\ncheck_frames <- check_frames[train_info, on = .(video_id)]\nhead(check_frames)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:58.684837Z","iopub.execute_input":"2025-10-03T20:34:58.686640Z","iopub.status.idle":"2025-10-03T20:34:58.740288Z","shell.execute_reply":"2025-10-03T20:34:58.738290Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"We can already see that the number of frames we can get from the video duration and FPS indicated in the train_csv (`csv_total_frames`) is different from that extracted form train_tracking files (`span_frames` = last_frame - first_frame + 1) \n\nLet's examine this difference:","metadata":{}},{"cell_type":"code","source":"check_frames[, diff_frames := span_frames - csv_total_frames]\ncheck_frames[, diff_time_sec := (span_frames / fps) - video_duration_sec]\n\nallowed_diff <- 100\nissues <- check_frames[abs(diff_frames) > allowed_diff]\n\nreport_issues(issues, \"Videos with number of frames missmatch (>100 frames):\")\nhead(issues[order(-abs(diff_frames))], 10)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:58.743210Z","iopub.execute_input":"2025-10-03T20:34:58.744817Z","iopub.status.idle":"2025-10-03T20:34:58.794248Z","shell.execute_reply":"2025-10-03T20:34:58.792186Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"quantile(issues[diff_frames > 0]$diff_frames)\nquantile(issues[diff_frames < 0]$diff_frames)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:58.797411Z","iopub.execute_input":"2025-10-03T20:34:58.798985Z","iopub.status.idle":"2025-10-03T20:34:58.828176Z","shell.execute_reply":"2025-10-03T20:34:58.826197Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"Check videos having annotations:","metadata":{}},{"cell_type":"code","source":"issues[lab_id %in% c(\"GroovyShrew\", \"PleasantMeerkat\")]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:58.830928Z","iopub.execute_input":"2025-10-03T20:34:58.832461Z","iopub.status.idle":"2025-10-03T20:34:58.860751Z","shell.execute_reply":"2025-10-03T20:34:58.858931Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"issues[diff_frames > 0, .N]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:58.863337Z","iopub.execute_input":"2025-10-03T20:34:58.864867Z","iopub.status.idle":"2025-10-03T20:34:58.881159Z","shell.execute_reply":"2025-10-03T20:34:58.879221Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"issues[diff_frames > 0, unique(lab_id)]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:58.883819Z","iopub.execute_input":"2025-10-03T20:34:58.885345Z","iopub.status.idle":"2025-10-03T20:34:58.901389Z","shell.execute_reply":"2025-10-03T20:34:58.899540Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"csv_total_frames and span_frames are not necessarily an error, e.g. it's expected if duration in train.csv is from real video metadataСoordinates and Arena size
","metadata":{}},{"cell_type":"markdown","source":"Ensure that min_x, max_x, min_y, and max_y from tracking and arena size do not exceed the video dimensions given in train.csv (video_width_pix, video_height_pix) (assuming the origin (0,0) is at the top-left corner of the video).","metadata":{}},{"cell_type":"code","source":"train_meta[, unique(arena_shape)]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:58.904250Z","iopub.execute_input":"2025-10-03T20:34:58.905750Z","iopub.status.idle":"2025-10-03T20:34:58.921834Z","shell.execute_reply":"2025-10-03T20:34:58.920049Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"check_coords <- track_summary[, .(video_id, min_x, max_x, min_y, max_y)]\ntmp <- train_meta[, .(\n video_id = as.character(video_id),\n lab_id,\n pix_per_cm_approx,\n video_width_pix,\n video_height_pix,\n arena_width_cm,\n arena_height_cm,\n arena_shape,\n tracking_method\n)]\ncheck_coords <- check_coords[tmp, on = .(video_id), nomatch = 0]\nhead(check_coords)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:58.924522Z","iopub.execute_input":"2025-10-03T20:34:58.926072Z","iopub.status.idle":"2025-10-03T20:34:58.983031Z","shell.execute_reply":"2025-10-03T20:34:58.981078Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Coordinates out of video","metadata":{}},{"cell_type":"code","source":"issues1 <- check_coords[(min_x < 0 | max_x > video_width_pix)]\nreport_issues(issues1, \"Videos with out of video dimentions x coordinate:\")\nhead(issues1[order(min_x), !c(\"min_y\", \"max_y\")], 10)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:58.985997Z","iopub.execute_input":"2025-10-03T20:34:58.987534Z","iopub.status.idle":"2025-10-03T20:34:59.027353Z","shell.execute_reply":"2025-10-03T20:34:59.024953Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"issues1[, unique(tracking_method)]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:59.030064Z","iopub.execute_input":"2025-10-03T20:34:59.031513Z","iopub.status.idle":"2025-10-03T20:34:59.053766Z","shell.execute_reply":"2025-10-03T20:34:59.051938Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"issues2 <- check_coords[(min_y < 0 | max_y > video_height_pix)]\nreport_issues(issues2, \"Videos with out of video dimentions y coordinate:\")\nhead(issues2[order(min_y), !c(\"min_x\", \"max_x\")], 10)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:59.056409Z","iopub.execute_input":"2025-10-03T20:34:59.057854Z","iopub.status.idle":"2025-10-03T20:34:59.098882Z","shell.execute_reply":"2025-10-03T20:34:59.097080Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"issues2[, unique(tracking_method)]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:59.101469Z","iopub.execute_input":"2025-10-03T20:34:59.102931Z","iopub.status.idle":"2025-10-03T20:34:59.119333Z","shell.execute_reply":"2025-10-03T20:34:59.117174Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"issues <- rbind(issues1, issues2)[, unique(.SD)]\nreport_issues(issues, \"Total videos with out of video dimentions coordinates:\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:59.122077Z","iopub.execute_input":"2025-10-03T20:34:59.123561Z","iopub.status.idle":"2025-10-03T20:34:59.142399Z","shell.execute_reply":"2025-10-03T20:34:59.140557Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"issues[!tracking_method %in% c(\"DeepLabCut\")]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-03T20:34:59.145154Z","iopub.execute_input":"2025-10-03T20:34:59.146690Z","iopub.status.idle":"2025-10-03T20:34:59.200106Z","shell.execute_reply":"2025-10-03T20:34:59.198215Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"Body parts count
","metadata":{}},{"cell_type":"markdown","source":"Ensures consistency between the body parts listed in the train.csv and the number of body parts in the tracking files.","metadata":{}},{"cell_type":"code","source":"split_body_parts <- function(x){\n x <- gsub('\"\"','', x, fixed = TRUE)\n x <- gsub(\"^\\\\[|\\\\]$\", \"\", x)\n trimws(strsplit(x, \", \", fixed = TRUE)[[1L]])\n}\nbp_check <- train_meta[, .(\n body_part = split_body_parts(body_parts_tracked)\n), by = .(lab_id, video_id)]\nbp_check[, video_id := as.character(video_id)]\n \nbp_check <- bp_check[\n , .(csv_num_bodyparts = uniqueN(body_part)),\n by = .(lab_id, video_id)\n]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T19:14:38.610446Z","iopub.execute_input":"2025-09-28T19:14:38.612385Z","iopub.status.idle":"2025-09-28T19:14:39.494693Z","shell.execute_reply":"2025-09-28T19:14:39.492812Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"bp_check <- bp_check[\n track_summary[, .(video_id, num_bodyparts)],\n on = .(video_id)\n] \nbp_check[csv_num_bodyparts != num_bodyparts]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T19:14:40.465318Z","iopub.execute_input":"2025-09-28T19:14:40.466871Z","iopub.status.idle":"2025-09-28T19:14:40.510506Z","shell.execute_reply":"2025-09-28T19:14:40.508821Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"Behaviors labeled
","metadata":{}},{"cell_type":"markdown","source":"Ensures consistency between the agent-target-action triplets listed in the train.csv (whitelist) and the annotated files.","metadata":{}},{"cell_type":"code","source":"clean_tokens <- function(x) {\n x <- gsub(\"\\\\\\\\\", \"\", as.character(x))\n x <- gsub(\"^['\\\"]+|['\\\"]+$\", \"\", x)\n x <- trimws(tolower(x))\n x[nchar(x) == 0] <- NA_character_\n x\n}\nnormalize_actor <- function(x) {\n x <- clean_tokens(x)\n is_mouse <- grepl(\"^mouse\\\\d+$\", x)\n out <- rep(NA_integer_, length(x))\n out[is_mouse] <- as.integer(sub(\"^mouse\", \"\", x[is_mouse]))\n out\n}\nparse_behaviors <- function(s) {\n s <- gsub('\"\"', '', s, fixed = TRUE)\n s <- gsub(\"^\\\\[|\\\\]$\", \"\", s)\n tokens <- trimws(unlist(strsplit(s, \",\", fixed = TRUE)))\n tokens <- tokens[nzchar(tokens)]\n stopifnot(length(tokens) %% 3 == 0)\n m <- matrix(tokens, ncol = 3, byrow = TRUE)\n dt <- as.data.table(m)\n setnames(dt, c(\"agent_raw\", \"target_raw\", \"action_raw\"))\n dt\n}","metadata":{"trusted":true,"_kg_hide-input":true,"execution":{"iopub.status.busy":"2025-09-28T19:14:43.291929Z","iopub.execute_input":"2025-09-28T19:14:43.293418Z","iopub.status.idle":"2025-09-28T19:14:43.309130Z","shell.execute_reply":"2025-09-28T19:14:43.307285Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"actions_check <- train_ann[, .(video_id, agent_id, target_id, action)][, unique(.SD)]\nhead(actions_check)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T19:14:45.801196Z","iopub.execute_input":"2025-09-28T19:14:45.802813Z","iopub.status.idle":"2025-09-28T19:14:45.837075Z","shell.execute_reply":"2025-09-28T19:14:45.834657Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"actions_check[, .(agent_id, target_id)][order(agent_id, target_id), unique(.SD)]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T19:14:48.322111Z","iopub.execute_input":"2025-09-28T19:14:48.323731Z","iopub.status.idle":"2025-09-28T19:14:48.353169Z","shell.execute_reply":"2025-09-28T19:14:48.350768Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"whitelist <- train_meta[\n behaviors_labeled != \"\",\n .(video_id = as.character(video_id), behaviors_labeled)\n]\nwhitelist <- whitelist[, {\n tmp <- parse_behaviors(behaviors_labeled)\n tmp[, `:=`(\n agent_id = normalize_actor(agent_raw),\n target_id = normalize_actor(target_raw),\n action = clean_tokens(action_raw)\n )]\n tmp[, target_id := ifelse(\n is.na(target_id) & clean_tokens(target_raw) == \"self\",\n agent_id, target_id\n )]\n}, by = video_id]\nhead(whitelist)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T19:14:51.002512Z","iopub.execute_input":"2025-09-28T19:14:51.004117Z","iopub.status.idle":"2025-09-28T19:14:52.666562Z","shell.execute_reply":"2025-09-28T19:14:52.664804Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"fsetdiff(\n unique(actions_check[, .(video_id, agent_id, target_id, action)]),\n unique(whitelist[, .(video_id, agent_id, target_id, action)])\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T19:16:20.417856Z","iopub.execute_input":"2025-09-28T19:16:20.419643Z","iopub.status.idle":"2025-09-28T19:16:20.449580Z","shell.execute_reply":"2025-09-28T19:16:20.447155Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"Actions duration
","metadata":{}},{"cell_type":"markdown","source":"Сhecks the consistency of annotated actions intervals:","metadata":{}},{"cell_type":"code","source":"actions_check <- copy(train_ann)\nby_cols <- c(\"agent_id\",\"target_id\", \"action\",\"video_id\")\nsetorderv(actions_check, c(by_cols, \"start_frame\", \"stop_frame\"))\nactions_check[, prev_stop := shift(stop_frame), by = by_cols]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T19:16:23.268797Z","iopub.execute_input":"2025-09-28T19:16:23.270492Z","iopub.status.idle":"2025-09-28T19:16:23.309720Z","shell.execute_reply":"2025-09-28T19:16:23.307846Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"Back-to-back intervals where one interval ends exactly when the next one starts:","metadata":{}},{"cell_type":"code","source":"actions_check[start_frame == prev_stop]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T19:16:25.704727Z","iopub.execute_input":"2025-09-28T19:16:25.706759Z","iopub.status.idle":"2025-09-28T19:16:25.732860Z","shell.execute_reply":"2025-09-28T19:16:25.731030Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"Overlapping intervals where the next interval starts before the previous one ended:","metadata":{}},{"cell_type":"code","source":"actions_check[start_frame < prev_stop]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T19:16:29.384788Z","iopub.execute_input":"2025-09-28T19:16:29.386397Z","iopub.status.idle":"2025-09-28T19:16:29.410496Z","shell.execute_reply":"2025-09-28T19:16:29.408680Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"Check actions duration:","metadata":{}},{"cell_type":"code","source":"actions_check[ , prev_stop := NULL]\nactions_check[, num_frames := stop_frame - start_frame + 1L]\nfps_dt <- train_meta[, .(lab_id, video_id = as.character(video_id), fps = frames_per_second)]\nactions_check[fps_dt, on = .(video_id), `:=` (fps = i.fps, lab_id = i.lab_id)]\n\nactions_check[, duration_sec := num_frames / fps]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T19:16:34.396284Z","iopub.execute_input":"2025-09-28T19:16:34.397902Z","iopub.status.idle":"2025-09-28T19:16:34.439604Z","shell.execute_reply":"2025-09-28T19:16:34.437723Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"issues <- actions_check[duration_sec <= 0.1][order(duration_sec)]\nreport_issues(issues, \"Actions with duration <= 0.1 seconds:\")\nhead(issues, 10)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T19:16:37.322474Z","iopub.execute_input":"2025-09-28T19:16:37.324107Z","iopub.status.idle":"2025-09-28T19:16:37.363888Z","shell.execute_reply":"2025-09-28T19:16:37.362104Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"actions_check[num_frames == 1][, unique(lab_id)]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T19:16:46.388047Z","iopub.execute_input":"2025-09-28T19:16:46.389668Z","iopub.status.idle":"2025-09-28T19:16:46.409485Z","shell.execute_reply":"2025-09-28T19:16:46.407162Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"Reciprocal interaction consistency
","metadata":{}},{"cell_type":"markdown","source":"In multi-agent behavior annotations, some actions are reciprocal by definition (e.g. “chase”: if mouse 1 chases mouse 2, then from mouse 2’s perspective the same event can be marked as “being chased”). Let's check if mirror records exists in this dataset and are the reciprocal actions and their span consistent.","metadata":{}},{"cell_type":"code","source":"actions_check <- train_ann[agent_id != target_id]\nactions_check[meta_files_dt, on = .(video_id), lab_id := i.lab_id]\nactions_check[, row_id := .I]\nhead(actions_check)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T20:01:52.170037Z","iopub.execute_input":"2025-09-28T20:01:52.171867Z","iopub.status.idle":"2025-09-28T20:01:52.223769Z","shell.execute_reply":"2025-09-28T20:01:52.221883Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"tolerance <- 20\n\ntic()\nres_list <- list()\nfor (idx in seq_len(nrow(actions_check))) {\n row_dt <- actions_check[row_id == idx]\n if (!nrow(row_dt)) next\n rec_dt <- actions_check[\n video_id == row_dt$video_id &\n agent_id == row_dt$target_id &\n target_id == row_dt$agent_id &\n start_frame >= row_dt$start_frame - tolerance &\n start_frame < row_dt$stop_frame &\n stop_frame <= row_dt$stop_frame + tolerance &\n stop_frame > row_dt$start_frame,\n .(video_id,\n row_id_rec = row_id,\n agent_id = target_id,\n target_id = agent_id,\n start_frame_rec = start_frame,\n stop_frame_rec = stop_frame,\n action_rec = action\n )\n ]\n if (nrow(rec_dt)) {\n actions_check <- actions_check[!row_id %in% rec_dt$row_id_rec]\n res_list[[idx]] <- merge(\n row_dt, rec_dt,\n by = c(\"video_id\", \"agent_id\",\"target_id\"),\n all = TRUE\n )\n }\n}\ntoc()\nreciprocal_actions <- rbindlist(res_list)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T20:01:54.392600Z","iopub.execute_input":"2025-09-28T20:01:54.394299Z","iopub.status.idle":"2025-09-28T20:08:37.384489Z","shell.execute_reply":"2025-09-28T20:08:37.382598Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"(Yes, a data.table join/antijoin would be 100 times faster. I just haven’t cracked it yet, until then: enjoy the brute-force loop.)","metadata":{}},{"cell_type":"code","source":"reciprocal_actions[, num_intersecting_frames := pmax(\n 0L,\n pmin(stop_frame, stop_frame_rec) -\n pmax(start_frame, start_frame_rec) + 1L\n )\n]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T20:10:39.710217Z","iopub.execute_input":"2025-09-28T20:10:39.711909Z","iopub.status.idle":"2025-09-28T20:10:39.727802Z","shell.execute_reply":"2025-09-28T20:10:39.725931Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"cat(\"Rows:\", reciprocal_actions[, .N])\nhead(reciprocal_actions)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-28T20:10:43.714342Z","iopub.execute_input":"2025-09-28T20:10:43.715928Z","iopub.status.idle":"2025-09-28T20:10:43.750880Z","shell.execute_reply":"2025-09-28T20:10:43.748747Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"