Mfu n'uche maka nchọpụta na-ezighi ezi
Mfufocal bụ ọrụ mfu gbanwetụrụ nke na-agbadata ihe atụ dị mfe ka onye nchọta nwee ike ilekwasị anya na ndị siri ike, ndị na-adịghị ahụkebe.
Nchịkọta
It solved the extreme background-versus-object imbalance that crippled one-stage object detectors.
Ime miri emi
N'ime nchọpụta ihe, onyonyo nwere ike ịnwe naanị ezigbo ihe ole na ole mana iri puku kwuru iri puku ebe ndị a na-achọ ntuli aka, ihe fọrọ nke nta ka ọ bụrụ ha niile dị mfe ndabere. Site na cross-entropy ọkọlọtọ, iju mmiri a dị mfe adịghị mma na-achịkwa gradient ma na-ewepụ ihe ndị na-adịghị ahụkebe. Mfufocal, ewepụtara na akwụkwọ RetinaNet 2017 nke Lin na ndị ọrụ ibe ya na Facebook AI, na-edozi nke a site n'ịba ụba cross-entropy site na ihe (1 - p_t) ^ gamma. Mgbe a na-ahazi ihe nlele na ntụkwasị obi yana nke ọma, p_t dị nso 1, yabụ ihe ahụ na-adaba na efu na ihe atụ nke nkewa nke ọma anaghị enye aka. Ọmụmaatụ siri ike, nke ezighi ezi na-edobe ibu dị nso. Na gamma gburugburu 2, RetinaNet dakọtara ma ọ bụ tie ihe nchọta ọkwa abụọ dị nwayọ dị ka Faster R-CNN ka ị na-anọ na netwọk otu ngafe dị mfe.
Nghọta nka nka
Ihe nlebara anya gamma na-achịkwa ka esi egbochi ihe atụ dị mfe ike: na gamma 0 focal loss hà nhata cross-entropy nkịtị, gamma dị elu na-eme ka elekwasị anya na ikpe siri ike. A na-ejikọta alfa na-edozi ahụ (mgbe 0.25 maka klaasị obere) na ya. N'ụzọ dị mkpa, ihe na-agbanwe agbanwe na-emegharị gradients, ọ bụghị naanị uru mfu, ya mere nkwado ndabere na-egosipụta n'ụzọ nkịtị na-emesi nlele anya na-enweghị mgbagha na-enweghị ngwungwu ihe atụ siri ike nke akwụkwọ ntuziaka.
Mmetụta atụmatụ
Ọnụ ego na mmefu ego
Mkpebi ihe owuwu ụlọ na-akwalite arụmọrụ yana ọnụ ahịa ọrụ ruo ọtụtụ afọ.
Mkpebi doro anya
Nkà mmụta nka na-enyere ndị otu egwuregwu aka ịhọrọ nchịkọta ziri ezi, ọ bụghị naanị nke kachasị ọhụrụ.
Quality akara
Nhọrọ injinia ka mma na-ebelata ihe omume ntụkwasị obi na mmepụta.
Ọdịnihu nke mfufocal maka nchọpụta na-ezighi ezi
Mfufocal abụrụla ngwa ndabara karịa RetinaNet, na-apụta na ihe nchọta dị ka FCOS, na ngalaba, yana na nhazi ogologo ọdụ. Ụdị dị iche iche dị ka mfu focal quality, mfu nkesa nkesa, na mfu varifocal na-emecha ya maka ihe nchọpụta na-enweghị arịlịka ọgbara ọhụrụ na nke dabere na mgbanwe. Ka nchọpụta na-atụgharị gaa n'ụdị amụma nhazi dị ka DETR nke na-eji matching bipartite, ụdị reweighting n'ụdị na-anọgide na-abụ ngwa bara uru ebe ọ bụla klaasị na-agbagọ nke ukwuu.
Mmejuputa n'ezie n'ụwa
Ịchọta obere akara okporo ụzọ ma ọ bụ ndị na-agafe agafe dị anya n'okirikiri ịnya ụgbọ ala kwụụrụ onwe ya ebe ọtụtụ pikselụ dị n'azụ.
Ịchọta etuto ma ọ bụ ọnya na-adịghị ahụkebe na nyocha ahụike nke anụ ahụ siri ike na-achịkwa.
Ịhụ ntụpọ n'ahịrị a na-emepụta ebe ihe ka ukwuu n'ime akụkụ ndị a na-enyocha bụ ihe nkịtị.
Ịmata obere ụgbọ mmiri ma ọ bụ ụgbọ ala na nnukwu satịlaịtị na onyonyo ikuku.
Ihe ize ndụ & okporo ụzọ nche
Ịkwalite otu akara ngosi nwere ike zoo adịghị ike sistemụ sara mbara.
A na-eledakarị ihe akụrụngwa na ụgwọ ọrụ anya.
Ọdịiche nchekwa na nleba anya nwere ike itolite ka sistemu na-adịwanye mgbagwoju anya.
Map mmejuputa
Kọwaa latency, ịdịmma na ebumnuche ọnụ ahịa tupu mmejuputa ya.
Benchmark n'okpuru ibu dị adị na ọnọdụ data.
Nleba anya akụrụngwa maka mperi, ịkpafu na mmetụta onye ọrụ.
Kwadebe ụzọ nzaghachi azụghachi azụ na ihe omume tupu ịchachaa.
Nọgide na-eme nchọpụta
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Focal Loss for Imbalanced Detection quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Ntuziaka na-esote
Netwọk Siamese na mfu Triplet
Ajụjụ a na-ajụkarị
What is Focal Loss for Imbalanced Detection?
Mfufocal bụ ọrụ mfu gbanwetụrụ nke na-agbadata ihe atụ dị mfe ka onye nchọta nwee ike ilekwasị anya na ndị siri ike, ndị na-adịghị ahụkebe. Ọ doziri oke n'okirikiri-na-ihe ahaghị nhata nke gwụchara ihe nchọta ihe nwere otu ọkwa.
Kedu isi nsogbu dị na nchọpụta ihe nwere otu ọkwa e mere iji lebara ya anya?
N'ime nchọpụta siri ike, ihe atụ n'okirikiri dị mfe karịrị ezigbo ihe ma na-achịkwa gradient ọkọlọtọ cross-entropy. Focal loss reweights dị ụkọ, ihe atụ siri ike na-agụ karịa.
Gịnị bụ ihe na-agbanwe agbanwe nke na-agbaba n'ihe efu na-abawanye na cross-entropy?
Nkwụsị uche bụ -(1 - p_t)^gamma * log(p_t). Okwu gamma (1 - p_t)^ gamma na-adaba na efu maka ibu amụma n'atụghị egwu.
Kedu ihe na-eme mfu n'uche mgbe gamma na-elekwasị anya na 0?
Na gamma = 0 ihe kpatara ya (1 - p_t) ^ 0 ha nhata 1, yabụ mfu focal na-ebelata kpọmkwem na ọkọlọtọ cross-entropy.
Kedu ihe gamma na-abawanye na-eme ka ọ bụrụ ihe atụ dị mfe na nkewa nke ọma?
Gamma dị elu na-eme ka okwu gamma (1 - p_t)^gamma na-ada ngwa ngwa maka p_t dị elu, yabụ ọmụmaatụ dị mfe na-enye aka ọbụnadị obere na mfu na gradient.
Kedu ụlọ ihe nchọta webatara mfufocal?
Ewebatara mfufocal na akwụkwọ 2017 RetinaNet, na-enyere onye na-achọpụta ihe otu ọkwa aka imeri ụzọ abụọ ziri ezi.