バッタ

【種牡馬通信簿（2025年末版）】 恒例のやつ ・コントレイルはどう評価すればよいの？ ・ナダル２年目はどうよ？ ・エピファネイアって23産が大活躍しているけど買いなの？ その疑問に答えが出るかも！？ 詳しくはブログにて。ひとまず結論の表だけ置いておきます。 ameblo.jp/xenogs2491/ent…

@k_mauvette モルジアナは「モルジアナ」じゃない方が走る説あるから笑

@drosshopper01 モルジアナじゃないガラベイヤ 流石に重賞じゃ厳しか。。

久々にこの記事読んだけど、突然歴史が始まるところと、最後に「内なるモルジアナ」とかいう謎の概念出てくるところ好き。 #S06" target="_blank" rel="nofollow noopener">ameblo.jp/xenogs2491/ent…

@mnch_pic 絶対深夜に書いてる笑

@drosshopper01 この熱感めちゃくちゃ好きすぎます🤣

@Lumineon_poke 追加募集で期待しましょう！

@drosshopper01 今年はここのページがないのが残念すぎます😢

サンデー、社台は6/8に測尺か。 noteはなるべく6/8当日更新にしたいと思いますが遅れたらごめんなさい。

クインズムーンの24（非）はデビュー早そうだな。 この仔、ツアーのときあんまり人気なくて、周りにあまり人いなかったんだけど、こっちが近づくと撫でてくれって頭寄せてくるの可愛すぎたから、ずっと更新追ってる。頑張ってほしい。

I’m really grateful that you’ve taken the time to think about this for me. Your concerns are completely reasonable. That said, I do have some ideas of my own regarding the two points you raised. 1. Given my objective, I may not need an extremely large dataset It is true that if the goal were to identify a tiny number of elite horses from the entire population, an enormous amount of data would be necessary. However, in Japan, horse racing is centrally managed by a national organization, the JRA, and the class system is very clearly defined. Specifically, the classes are: Maiden 1-win class 2-win class 3-win class Open class My goal is not to pinpoint future Grade race winners within the Open class. Rather, it would be sufficient to statistically demonstrate that as the model’s score increases, the proportion of horses reaching higher classes also increases. For that kind of objective, I believe the required amount of data may not be as large as one might initially assume. 2. Japan has a substantial number of high-quality walk videos In Japan, there is a well-developed system called hitokuchi banushi, where ownership of a racehorse is divided into 40 to 500 shares. There are around 15 organizations that operate this kind of shared ownership system. Among them, Sunday Racing, Silk Horse Club, and Carrot Club are all closely connected to Northern Farm, and Shadai Thoroughbred Club is also operated by relatives of the same group. These clubs provide walk videos of their horses at the time of recruitment. Because these clubs are run by closely related organizations, the videos are not perfectly standardized, but they are generally filmed from similar angles. Unfortunately, the camera distance is not always consistent. Still, the number of videos is about 300 per year. If I collect data over 10 years, that would amount to around 3,000 horses, which I believe should be enough for the objective described in point 1. Even better, these horses are provided with height data along with the videos. I expect that using this height data could help address the scale-difference problem to some extent.

Not a problem. Glad to help Also, if your goal is to separate the foals that turn out to be elite vs non-elite, two other concerns: - You need a lot of data to run inference on to build a predictive model. The positive case (elite) is small (~3%) and you’re better off modeling as a binary outcome with even groups (elite vs really slow) and you need ~500 elite so there needs to be ~16,000 videos to choose from, and you use 1000 in your model (10 fold cross validation) otherwise you’re underpowered. - in all these videos the distance from the camera to the horse is unknown so any measurements are not to scale. Even if you augment with monocular depth estimation like Depth Pro or Depth Anything V3 you still won’t get accurate ‘to scale’ measurements of the horses. This is a fatal flaw of using online videos as the size/body length of the horse has a direct correlation to stride length and cadence on the racetrack. The lack of knowledge of actual size with these videos is a large source of error in prediction.

ついに、、、あとはデータを集めるだけ。

@dddiiisss444666 ご助言ありがとうございます！ 色々な馬の色々なパターン混ぜてみようと思います！

@drosshopper01 私も彼の言う数ほどフレームをこなしました。 体感で白斑のせいでブレる気がしましたので 適宜、白斑のない馬を混ぜるのが良さそうです。

海外ニキのありがたいアドバイス

@rano_hoshipan おそらくですが、動画の背景は定点をブレさせる要因になると思います。 ただ、海外の成功例を見ていると、十分に訓練すればそれは解決しそうです

@drosshopper01 なるほどー。動画の背景も多少関係がありそうな

Thank you so much. There is nothing more valuable than receiving advice from someone who has already gone through this process. I wasn’t sure how many videos I should use, what kinds of videos would be appropriate, or how many frames I should label for training, so your advice was extremely helpful.

My advice. - get a representation of videos in different circumstances but make sure they are somewhat even in count. So get left to right walk video of 5 bay, 5 black, 5 chestnut and 5 grey foals (20 total) It will generalize better - do the k-means image extraction across each of the 20 videos. 30 frames will be enough per video (600 frames total). - label them consistently and then train for a decent amount of epochs. - review the ones with large errors. You will usually find it’s where you’ve placed a marker incorrectly or the landmark isn’t clear anyway and should be removed. Retrain. - even with this you’ll find that on all videos there is a sliding window where the landmarks are placed more accurately. The frames just before and just after the camera is perpendicular to the horse are usually best. You can usually get a full stride cycle of good data to use in downstream analysis.

@umaaaaanta ありがとうございます！

@drosshopper01 ノーザン馬当歳なら2022年以降、馬体重のみ公開されています。1歳馬も馬体重はノーザン馬しか公開されていないと思います。 northernfarm.jp/sale/select/20…

あー、セレクトでも測尺は１歳馬しか公表されないのか。残念。 jrha.or.jp/selectsale/sal…

@ccmdu2GbFdNBUcG アイちゃん牝系なのにねー。まあそういう謳い文句でセールで高く売れそうではある。実際去年、25産が父レイデオロ牡馬で7920万円でミックスセールで売れてるし。

@drosshopper01 ユタライトみたいな仔を今後も産むなら勿体無い 血統的にも

ユナカイトはもう完全によその子になってしまったのかな。２勝馬だから仕方ないけど。

当歳馬の体重予測作れば当歳のセレクトセールの評価もできるんか。今なら4,5時間でできそう笑 セレクトの過去の測尺ってどこにあるんだっけ？てか、のびしろ君作る時、クソ大変な思いしてサンデー、シルク、キャロットの測尺データ集めたけど、セレクトのデータ使えばもっと簡単にできたのでは。

なんだよ！！あんまりいねぇじゃねえか！！ それシルクのだろ！！って言わせろよ！！！ まあ、あえて言うなら25産はクールキャットぐらい？まあでもレイクヴィラ産は仔がシルクで回ってこないことがほとんどだし、父ブリモルだし、出資者は残念だろうけど、まあしゃあないかという感じはある。

あ、そういえばセレクトセールのリストに、それシルクのだろ！！って文句つける儀式まだやってなかった。

これやばい。走ってるのに各定点が全くブレてないの震えるわ。

@aki_sawayaka バッタびっくり(・o・)

@drosshopper01 そりゃ大反響でしょーに！

想像の10倍ぐらい反響があって驚いてます。 分かる人には分かる映像として、一部の人にニヤッとしてもらえればと思ってアップましたが、想像よりも多くの人が馬体・歩様の客観データによる分析を待っていることがわかり、モチベが更に上がりました。 シルクに間に合ってくれーー！